diff --git a/.gitignore b/.gitignore
index d6ae9214..3bb04cb2 100644
--- a/.gitignore
+++ b/.gitignore
@@ -33,3 +33,6 @@ dist-ssr
 # other secret files
 .env.local
 .env.langfuse
+
+# local models config
+local-models.json
\ No newline at end of file
diff --git a/README.md b/README.md
index 00a10bd3..72d069ba 100644
--- a/README.md
+++ b/README.md
@@ -211,6 +211,53 @@ npm run dev
   NGROK_URL="<NGROK URL>" # Optional local Supabase Storage tunnel for provider-readable signed URLs
   ```
 
+### 3. Optional Local LLM Models
+
+To add your own local models (LM Studio, Ollama, LiteLLM, etc.) alongside the cloud models:
+
+1. Start one or more local OpenAI-compatible servers (for example LM Studio on port `1234`).
+2. Copy the example catalog and edit entries to match your servers:
+
+   ```bash
+   cp local-models.json.example local-models.json
+   ```
+
+   Each entry needs an `id`, `name`, `description`, and `baseUrl`. The `id` must match what your server exposes.
+
+   Example:
+
+   ```json
+   {
+     "id": "qwen2.5-vl-7b-instruct",
+     "name": "Qwen2.5 VL 7B",
+     "description": "Local VLM",
+     "baseUrl": "http://localhost:1234",
+     "supportsTools": true,
+     "supportsVision": true,
+     "useForAux": true
+   },
+   {
+     "id": "deepseek-v4-flash",
+     "name": "DeepSeek v4 Flash",
+     "description": "Third Party LLM",
+     "baseUrl": "https://api.deepseek.com",
+     "apiKey": "DEEPSEEK_API_KEY",
+     "supportsTools": true,
+     "supportsThinking": true,
+     "supportsVision": false,
+     "supportsForcedToolChoice": false,
+     "useForAux": false
+   }
+   ```
+
+3. Set local model API keys in `.env.local`.
+   - `apiKey` in `local-models.json` should be the env var name, not the secret value.
+     - Example: `"apiKey": "DEEPSEEK_API_KEY"` and then set `DEEPSEEK_API_KEY="<your key>"` in `.env.local`.
+
+4. Restart `npm run dev`. Models with a `baseUrl` appear in the parametric model picker.
+
+Cloud models (Gemini, Claude, OpenAI, Z-AI) keep using their existing cloud routes. Only ids listed in `local-models.json` with a `baseUrl` are sent to that model's server.
+
 ## 🌐 Setting Up ngrok for Local Development
 
 CADAM uses public URLs for provider callbacks and local signed storage URLs:
diff --git a/local-models.json.example b/local-models.json.example
new file mode 100644
index 00000000..d6f834aa
--- /dev/null
+++ b/local-models.json.example
@@ -0,0 +1,24 @@
+[
+  {
+    "id": "qwen2.5-vl-7b-instruct",
+    "name": "Qwen2.5 VL 7B",
+    "description": "Local VLM",
+    "baseUrl": "http://localhost:1234",
+    "supportsTools": true,
+    "supportsThinking": false,
+    "supportsVision": true,
+    "useForAux": true
+  },
+  {
+    "id": "deepseek-v4-flash",
+    "name": "DeepSeek v4 Flash",
+    "description": "Third Party LLM",
+    "baseUrl": "https://api.deepseek.com",
+    "apiKey": "DEEPSEEK_API_KEY",
+    "supportsTools": true,
+    "supportsThinking": true,
+    "supportsVision": false,
+    "supportsForcedToolChoice": false,
+    "useForAux": false
+  }
+]
diff --git a/package-lock.json b/package-lock.json
index 7e96a05f..11f41c1b 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -365,6 +365,7 @@
       "resolved": "https://registry.npmjs.org/@babel/core/-/core-7.29.7.tgz",
       "integrity": "sha512-RgHBCvtjbOK2gXSNBNIkNoEc9qoVEtau3hj8gEqKQuL3HZAibKarWFEI3Lfm6EYKkLalOh8eSrj9b+ch9H/VBA==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@babel/code-frame": "^7.29.7",
         "@babel/generator": "^7.29.7",
@@ -672,33 +673,10 @@
       "integrity": "sha512-U+HFai5+zmJCkK86QsaJtoITlboZHBqrVketcO2ROv865xfCMSFpELQoz1GkX5GzME8pTa+3kbKrZHQtI0gdbw==",
       "license": "Apache-2.0"
     },
-    "node_modules/@emnapi/core": {
-      "version": "1.10.0",
-      "resolved": "https://registry.npmjs.org/@emnapi/core/-/core-1.10.0.tgz",
-      "integrity": "sha512-yq6OkJ4p82CAfPl0u9mQebQHKPJkY7WrIuk205cTYnYe+k2Z8YBh11FrbRG/H6ihirqcacOgl2BIO8oyMQLeXw==",
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "dependencies": {
-        "@emnapi/wasi-threads": "1.2.1",
-        "tslib": "^2.4.0"
-      }
-    },
-    "node_modules/@emnapi/runtime": {
-      "version": "1.10.0",
-      "resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.10.0.tgz",
-      "integrity": "sha512-ewvYlk86xUoGI0zQRNq/mC+16R1QeDlKQy21Ki3oSYXNgLb45GV1P6A0M+/s6nyCuNDqe5VpaY84BzXGwVbwFA==",
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "dependencies": {
-        "tslib": "^2.4.0"
-      }
-    },
     "node_modules/@emnapi/wasi-threads": {
-      "version": "1.2.1",
-      "resolved": "https://registry.npmjs.org/@emnapi/wasi-threads/-/wasi-threads-1.2.1.tgz",
-      "integrity": "sha512-uTII7OYF+/Mes/MrcIOYp5yOtSMLBWSIoLPpcgwipoiKbli6k322tcoFsxoIIxPDqW01SQGAgko4EzZi2BNv2w==",
+      "version": "1.2.2",
+      "resolved": "https://registry.npmjs.org/@emnapi/wasi-threads/-/wasi-threads-1.2.2.tgz",
+      "integrity": "sha512-c95qOXkHdydNKhscBTebqEC1CVAZpyqOfVfBzQ1qgzyl3gfeldUjIggDbIZgDKsHLgnsM+igH7TJ/eAasaVuMA==",
       "dev": true,
       "license": "MIT",
       "optional": true,
@@ -2763,6 +2741,7 @@
       "version": "9.2.0",
       "resolved": "https://registry.npmjs.org/@react-three/fiber/-/fiber-9.2.0.tgz",
       "integrity": "sha512-esZe+E9T/aYEM4HlBkirr/yRE8qWTp9WUsLISyHHMCHKlJv85uc5N4wwKw+Ay0QeTSITw6T9Q3Svpu383Q+CSQ==",
+      "peer": true,
       "dependencies": {
         "@babel/runtime": "^7.17.8",
         "@types/react-reconciler": "^0.28.9",
@@ -3031,6 +3010,40 @@
         "node": "^20.19.0 || >=22.12.0"
       }
     },
+    "node_modules/@rolldown/binding-wasm32-wasi/node_modules/@emnapi/core": {
+      "version": "1.10.0",
+      "resolved": "https://registry.npmjs.org/@emnapi/core/-/core-1.10.0.tgz",
+      "integrity": "sha512-yq6OkJ4p82CAfPl0u9mQebQHKPJkY7WrIuk205cTYnYe+k2Z8YBh11FrbRG/H6ihirqcacOgl2BIO8oyMQLeXw==",
+      "dev": true,
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "@emnapi/wasi-threads": "1.2.1",
+        "tslib": "^2.4.0"
+      }
+    },
+    "node_modules/@rolldown/binding-wasm32-wasi/node_modules/@emnapi/runtime": {
+      "version": "1.10.0",
+      "resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.10.0.tgz",
+      "integrity": "sha512-ewvYlk86xUoGI0zQRNq/mC+16R1QeDlKQy21Ki3oSYXNgLb45GV1P6A0M+/s6nyCuNDqe5VpaY84BzXGwVbwFA==",
+      "dev": true,
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "tslib": "^2.4.0"
+      }
+    },
+    "node_modules/@rolldown/binding-wasm32-wasi/node_modules/@emnapi/wasi-threads": {
+      "version": "1.2.1",
+      "resolved": "https://registry.npmjs.org/@emnapi/wasi-threads/-/wasi-threads-1.2.1.tgz",
+      "integrity": "sha512-uTII7OYF+/Mes/MrcIOYp5yOtSMLBWSIoLPpcgwipoiKbli6k322tcoFsxoIIxPDqW01SQGAgko4EzZi2BNv2w==",
+      "dev": true,
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "tslib": "^2.4.0"
+      }
+    },
     "node_modules/@rolldown/binding-win32-arm64-msvc": {
       "version": "1.0.3",
       "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-arm64-msvc/-/binding-win32-arm64-msvc-1.0.3.tgz",
@@ -4581,6 +4594,7 @@
       "resolved": "https://registry.npmjs.org/@types/mdast/-/mdast-4.0.4.tgz",
       "integrity": "sha512-kGaNbPh1k7AFzgpud/gMdvIm5xuECykRR+JnWKQno9TAXVa6WIVCGTPvYGekIDL4uwCZQSYbUxNBSb1aUo79oA==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@types/unist": "*"
       }
@@ -4615,6 +4629,7 @@
       "resolved": "https://registry.npmjs.org/@types/react/-/react-19.2.1.tgz",
       "integrity": "sha512-1U5NQWh/GylZQ50ZMnnPjkYHEaGhg6t5i/KI0LDDh3t4E3h3T3vzm+GLY2BRzMfIjSBwzm6tginoZl5z0O/qsA==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "csstype": "^3.0.2"
       }
@@ -4625,6 +4640,7 @@
       "integrity": "sha512-/EEvYBdT3BflCWvTMO7YkYBHVE9Ci6XdqZciZANQgKpaiDRGOLIlRo91jbTNRQjgPFWVaRxcYc0luVNFitz57A==",
       "devOptional": true,
       "license": "MIT",
+      "peer": true,
       "peerDependencies": {
         "@types/react": "^19.2.0"
       }
@@ -4652,6 +4668,7 @@
       "version": "0.160.0",
       "resolved": "https://registry.npmjs.org/@types/three/-/three-0.160.0.tgz",
       "integrity": "sha512-jWlbUBovicUKaOYxzgkLlhkiEQJkhCVvg4W2IYD2trqD2om3VK4DGLpHH5zQHNr7RweZK/5re/4IVhbhvxbV9w==",
+      "peer": true,
       "dependencies": {
         "@types/stats.js": "*",
         "@types/webxr": "*",
@@ -4729,6 +4746,7 @@
       "integrity": "sha512-Zhy8HCvBUEfBECzIl1PKqF4p11+d0aUJS1GeUiuqK9WmOug8YCmC4h4bjyBvMyAMI9sbRczmrYL5lKg/YMbrcQ==",
       "dev": true,
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@typescript-eslint/scope-manager": "8.38.0",
         "@typescript-eslint/types": "8.38.0",
@@ -5039,6 +5057,7 @@
       "resolved": "https://registry.npmjs.org/acorn/-/acorn-8.16.0.tgz",
       "integrity": "sha512-UVJyE9MttOsBQIDKw1skb9nAwQuR5wuGD3+82K6JgJlm/Y+KI92oNsMNGZCYdDsVtRHSak0pcV5Dno5+4jh9sw==",
       "license": "MIT",
+      "peer": true,
       "bin": {
         "acorn": "bin/acorn"
       },
@@ -5072,6 +5091,7 @@
       "resolved": "https://registry.npmjs.org/ai/-/ai-6.0.177.tgz",
       "integrity": "sha512-1xQtbeWwNcLyyM86ixZhkKvT+WRXc1lvarIKqPVtsyn8F9NDikwUMBqYu+aQKDgMht50SMXh4qboYuU8MeHZZA==",
       "license": "Apache-2.0",
+      "peer": true,
       "dependencies": {
         "@ai-sdk/gateway": "3.0.112",
         "@ai-sdk/provider": "3.0.10",
@@ -5345,6 +5365,7 @@
           "url": "https://github.com/sponsors/ai"
         }
       ],
+      "peer": true,
       "dependencies": {
         "caniuse-lite": "^1.0.30001726",
         "electron-to-chromium": "^1.5.173",
@@ -5799,6 +5820,7 @@
       "integrity": "sha512-wUR89x/Rw7/8t+vn0CmGDYM9TD6VtARGb0LD5jq2wjtMy1vCP4M+sm6N6TigWeTYvnA8MoW29NqqXD0ep0rfBA==",
       "devOptional": true,
       "license": "MIT",
+      "peer": true,
       "peerDependencies": {
         "srvx": ">=0.11.5"
       },
@@ -5864,6 +5886,7 @@
       "resolved": "https://registry.npmjs.org/cytoscape/-/cytoscape-3.33.2.tgz",
       "integrity": "sha512-sj4HXd3DokGhzZAdjDejGvTPLqlt84vNFN8m7bGsOzDY5DyVcxIb2ejIXat2Iy7HxWhdT/N1oKyheJ5YdpsGuw==",
       "license": "MIT",
+      "peer": true,
       "engines": {
         "node": ">=0.10"
       }
@@ -6266,6 +6289,7 @@
       "resolved": "https://registry.npmjs.org/d3-selection/-/d3-selection-3.0.0.tgz",
       "integrity": "sha512-fmTRWbNMmsmWq6xJV8D19U/gw/bwrHfNXxrIN+HfZgnzqTHp9jOmKMhsTUjXOJnZOdZY9Q28y4yebKzqDKlxlQ==",
       "license": "ISC",
+      "peer": true,
       "engines": {
         "node": ">=12"
       }
@@ -6616,7 +6640,8 @@
     "node_modules/embla-carousel": {
       "version": "8.6.0",
       "resolved": "https://registry.npmjs.org/embla-carousel/-/embla-carousel-8.6.0.tgz",
-      "integrity": "sha512-SjWyZBHJPbqxHOzckOfo8lHisEaJWmwd23XppYFYVh10bU66/Pn5tkVkbkCMZVdbUE5eTCI2nD8OyIP4Z+uwkA=="
+      "integrity": "sha512-SjWyZBHJPbqxHOzckOfo8lHisEaJWmwd23XppYFYVh10bU66/Pn5tkVkbkCMZVdbUE5eTCI2nD8OyIP4Z+uwkA==",
+      "peer": true
     },
     "node_modules/embla-carousel-react": {
       "version": "8.6.0",
@@ -6744,6 +6769,7 @@
       "resolved": "https://registry.npmjs.org/eslint/-/eslint-9.31.0.tgz",
       "integrity": "sha512-QldCVh/ztyKJJZLr4jXNUByx3gR+TDYZCRXEktiZoUR3PGy4qCmSbkxcIle8GEwGpb5JBZazlaJ/CxLidXdEbQ==",
       "dev": true,
+      "peer": true,
       "dependencies": {
         "@eslint-community/eslint-utils": "^4.2.0",
         "@eslint-community/regexpp": "^4.12.1",
@@ -9420,6 +9446,7 @@
         }
       ],
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@types/debug": "^4.0.0",
         "debug": "^4.0.0",
@@ -10056,7 +10083,8 @@
           "url": "https://opencollective.com/unified"
         }
       ],
-      "license": "MIT"
+      "license": "MIT",
+      "peer": true
     },
     "node_modules/micromatch": {
       "version": "4.0.8",
@@ -10284,30 +10312,13 @@
         }
       }
     },
-    "node_modules/nitro/node_modules/chokidar": {
-      "version": "5.0.0",
-      "resolved": "https://registry.npmjs.org/chokidar/-/chokidar-5.0.0.tgz",
-      "integrity": "sha512-TQMmc3w+5AxjpL8iIiwebF73dRDF4fBIieAqGn9RGCWaEVwQ6Fb2cGe31Yns0RRIzii5goJ1Y7xbMwo1TxMplw==",
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "peer": true,
-      "dependencies": {
-        "readdirp": "^5.0.0"
-      },
-      "engines": {
-        "node": ">= 20.19.0"
-      },
-      "funding": {
-        "url": "https://paulmillr.com/funding/"
-      }
-    },
     "node_modules/nitro/node_modules/db0": {
       "version": "0.3.4",
       "resolved": "https://registry.npmjs.org/db0/-/db0-0.3.4.tgz",
       "integrity": "sha512-RiXXi4WaNzPTHEOu8UPQKMooIbqOEyqA1t7Z6MsdxSCeb8iUC9ko3LcmsLmeUt2SM5bctfArZKkRQggKZz7JNw==",
       "dev": true,
       "license": "MIT",
+      "peer": true,
       "peerDependencies": {
         "@electric-sql/pglite": "*",
         "@libsql/client": "*",
@@ -10337,18 +10348,6 @@
         }
       }
     },
-    "node_modules/nitro/node_modules/lru-cache": {
-      "version": "11.5.1",
-      "resolved": "https://registry.npmjs.org/lru-cache/-/lru-cache-11.5.1.tgz",
-      "integrity": "sha512-RPimw/7aMdv2oqRrxKwvZXcPfwBrn/JZ2xYcY9Hus/6LaS3VOAKVWKWgNLCFSiOm1ESXinjsDlidVU7JlnCN2A==",
-      "dev": true,
-      "license": "BlueOak-1.0.0",
-      "optional": true,
-      "peer": true,
-      "engines": {
-        "node": "20 || >=22"
-      }
-    },
     "node_modules/nitro/node_modules/readdirp": {
       "version": "5.0.0",
       "resolved": "https://registry.npmjs.org/readdirp/-/readdirp-5.0.0.tgz",
@@ -10356,7 +10355,6 @@
       "dev": true,
       "license": "MIT",
       "optional": true,
-      "peer": true,
       "engines": {
         "node": ">= 20.19.0"
       },
@@ -10578,7 +10576,8 @@
       "resolved": "https://registry.npmjs.org/ofetch/-/ofetch-2.0.0-alpha.3.tgz",
       "integrity": "sha512-zpYTCs2byOuft65vI3z43Dd6iSdFbOZZLb9/d21aCpx2rGastVU9dOCv0lu4ykc1Ur1anAYjDi3SUvR0vq50JA==",
       "dev": true,
-      "license": "MIT"
+      "license": "MIT",
+      "peer": true
     },
     "node_modules/ohash": {
       "version": "2.0.11",
@@ -10943,6 +10942,7 @@
         }
       ],
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "nanoid": "^3.3.12",
         "picocolors": "^1.1.1",
@@ -11108,6 +11108,7 @@
       "version": "3.6.2",
       "resolved": "https://registry.npmjs.org/prettier/-/prettier-3.6.2.tgz",
       "integrity": "sha512-I7AIg5boAr5R0FFtJ6rCfD+LFsWHp81dolrFD8S79U9tb8Az2nGrJncnMSnys+bpQJfRUzqs9hnA81OAA3hCuQ==",
+      "peer": true,
       "bin": {
         "prettier": "bin/prettier.cjs"
       },
@@ -11314,6 +11315,7 @@
       "resolved": "https://registry.npmjs.org/react/-/react-19.2.1.tgz",
       "integrity": "sha512-DGrYcCWK7tvYMnWh79yrPHt+vdx9tY+1gPZa7nJQtO/p8bLTDaHp4dzwEhQB7pZ4Xe3ok4XKuEPrVuc+wlpkmw==",
       "license": "MIT",
+      "peer": true,
       "engines": {
         "node": ">=0.10.0"
       }
@@ -11333,6 +11335,7 @@
       "resolved": "https://registry.npmjs.org/react-dom/-/react-dom-19.2.1.tgz",
       "integrity": "sha512-ibrK8llX2a4eOskq1mXKu/TGZj9qzomO+sNfO98M6d9zIPOEhlBkMkBUBLd1vgS0gQsLDBzA+8jJBVXDnfHmJg==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "scheduler": "^0.27.0"
       },
@@ -11364,6 +11367,7 @@
       "version": "7.60.0",
       "resolved": "https://registry.npmjs.org/react-hook-form/-/react-hook-form-7.60.0.tgz",
       "integrity": "sha512-SBrYOvMbDB7cV8ZfNpaiLcgjH/a1c7aK0lK+aNigpf4xWLO8q+o4tcvVurv3c4EOyzn/3dCsYt4GKD42VvJ/+A==",
+      "peer": true,
       "engines": {
         "node": ">=18.0.0"
       },
@@ -12031,6 +12035,7 @@
       "resolved": "https://registry.npmjs.org/seroval/-/seroval-1.5.4.tgz",
       "integrity": "sha512-46uFvgrXTVxZcUorgSSRZ4y+ieqLLQRMlG4bnCZKW3qI6BZm7Rg4ntMW4p1mILEEBZWrFlcpp0AyIIlM6jD9iw==",
       "license": "MIT",
+      "peer": true,
       "engines": {
         "node": ">=10"
       }
@@ -12557,6 +12562,7 @@
       "version": "3.4.17",
       "resolved": "https://registry.npmjs.org/tailwindcss/-/tailwindcss-3.4.17.tgz",
       "integrity": "sha512-w33E2aCvSDP0tW9RZuNXadXlkHXqFzSkQew/aIa2i/Sj8fThxwovwlXHSPXTbAHwEIhBFXAedUhP2tueAKP8Og==",
+      "peer": true,
       "dependencies": {
         "@alloc/quick-lru": "^5.2.0",
         "arg": "^5.0.2",
@@ -12628,7 +12634,8 @@
     "node_modules/three": {
       "version": "0.160.1",
       "resolved": "https://registry.npmjs.org/three/-/three-0.160.1.tgz",
-      "integrity": "sha512-Bgl2wPJypDOZ1stAxwfWAcJ0WQf7QzlptsxkjYiURPz+n5k4RBDLsq+6f9Y75TYxn6aHLcWz+JNmwTOXWrQTBQ=="
+      "integrity": "sha512-Bgl2wPJypDOZ1stAxwfWAcJ0WQf7QzlptsxkjYiURPz+n5k4RBDLsq+6f9Y75TYxn6aHLcWz+JNmwTOXWrQTBQ==",
+      "peer": true
     },
     "node_modules/three-mesh-bvh": {
       "version": "0.8.3",
@@ -12724,6 +12731,7 @@
       "resolved": "https://registry.npmjs.org/picomatch/-/picomatch-4.0.4.tgz",
       "integrity": "sha512-QP88BAKvMam/3NxH6vj2o21R6MjxZUAd6nlwAS/pnGvN9IVLocLHxGYIzFhg6fUQ+5th6P4dv4eW9jX3DSIj7A==",
       "license": "MIT",
+      "peer": true,
       "engines": {
         "node": ">=12"
       },
@@ -12902,6 +12910,7 @@
       "integrity": "sha512-p1diW6TqL9L07nNxvRMM7hMMw4c5XOo/1ibL4aAIGmSAt9slTE1Xgw5KWuof2uTOvCg9BY7ZRi+GaF+7sfgPeQ==",
       "dev": true,
       "license": "Apache-2.0",
+      "peer": true,
       "bin": {
         "tsc": "bin/tsc",
         "tsserver": "bin/tsserver"
@@ -12980,6 +12989,7 @@
       "resolved": "https://registry.npmjs.org/unified/-/unified-11.0.5.tgz",
       "integrity": "sha512-xKvGhPWw3k84Qjh8bI3ZeJjqnyadK+GEFtazSfZv/rKeTkTjOJho6mFqh2SM96iIcZokxiOpg78GazTSg8+KHA==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@types/unist": "^3.0.0",
         "bail": "^2.0.0",
@@ -13297,6 +13307,7 @@
       "integrity": "sha512-h9bXPmJichP5fLmVQo3PyaGSDE2n3aPuomeAlVRm0JLmt4rY6zmPKd59HYI4LNW8oTK7tlTsuC7l/m7awx9Jcw==",
       "devOptional": true,
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "lightningcss": "^1.32.0",
         "picomatch": "^4.0.4",
@@ -13655,6 +13666,7 @@
       "resolved": "https://registry.npmjs.org/zod/-/zod-3.25.76.tgz",
       "integrity": "sha512-gzUt/qt81nXsFGKIFcC3YnfEAx5NkunCfnDlvuBSSFS02bcXu4Lmea0AFIUwbLWxWPx3d9p8S5QoaujKcNQxcQ==",
       "license": "MIT",
+      "peer": true,
       "funding": {
         "url": "https://github.com/sponsors/colinhacks"
       }
diff --git a/shared/imageRefs.ts b/shared/imageRefs.ts
index 26cef4ed..e9f90dd4 100644
--- a/shared/imageRefs.ts
+++ b/shared/imageRefs.ts
@@ -29,6 +29,14 @@ export function imageStoragePath(
   return `${userId}/${conversationId}/${imageId}`;
 }
 
+export function inspectionPreviewStoragePath(
+  userId: string,
+  conversationId: string,
+  toolCallId: string,
+): string {
+  return `${userId}/${conversationId}/inspection-preview-${toolCallId}`;
+}
+
 // Canonical reference persisted in a file part's `url`. Kept in the exact
 // shape the 2026-05-18 AI-SDK backfill produced so every row in
 // `messages.parts` carries a single, uniform format. Never fetched directly —
diff --git a/shared/localModels.test.ts b/shared/localModels.test.ts
new file mode 100644
index 00000000..750e1076
--- /dev/null
+++ b/shared/localModels.test.ts
@@ -0,0 +1,339 @@
+import assert from 'node:assert/strict';
+import { describe, it } from 'node:test';
+import {
+  getAuxiliaryLocalModel,
+  getLocalModels,
+  isMissingLocalBaseUrl,
+  isLocalModelId,
+  localModelToPickerConfig,
+  normalizeOpenRouterBaseUrl,
+  parseLocalModelsJson,
+  providerFor,
+} from './localModels.ts';
+
+const sampleLocalModels = parseLocalModelsJson([
+  {
+    id: 'qwen2.5-coder-7b',
+    name: 'Qwen',
+    description: 'Local',
+    baseUrl: 'http://localhost:1234',
+  },
+]);
+
+describe('parseLocalModelsJson', () => {
+  it('returns an empty array for non-array input', () => {
+    assert.deepEqual(parseLocalModelsJson(null), []);
+    assert.deepEqual(parseLocalModelsJson({}), []);
+  });
+
+  it('parses valid entries and skips invalid ones', () => {
+    assert.deepEqual(
+      parseLocalModelsJson([
+        {
+          id: 'qwen2.5-coder-7b',
+          name: 'Qwen',
+          description: 'Local coder',
+          apiKey: 'TEST_ENV_KEY_NAME',
+          supportsTools: true,
+        },
+        { id: '', name: 'Bad', description: 'x' },
+      ]),
+      [
+        {
+          id: 'qwen2.5-coder-7b',
+          name: 'Qwen',
+          description: 'Local coder',
+          apiKey: 'TEST_ENV_KEY_NAME',
+          supportsTools: true,
+        },
+      ],
+    );
+  });
+
+  it('accepts entries without apiKey', () => {
+    assert.deepEqual(
+      parseLocalModelsJson([
+        {
+          id: 'qwen3-vl-30b',
+          name: 'Qwen3 VL 30B',
+          description: 'Local VLM',
+          baseUrl: 'http://localhost:8004/v1',
+          supportsTools: true,
+        },
+      ]),
+      [
+        {
+          id: 'qwen3-vl-30b',
+          name: 'Qwen3 VL 30B',
+          description: 'Local VLM',
+          baseUrl: 'http://localhost:8004/v1',
+          supportsTools: true,
+        },
+      ],
+    );
+  });
+
+  it('keeps the first entry when duplicate ids appear', () => {
+    assert.deepEqual(
+      parseLocalModelsJson([
+        {
+          id: 'dup-model',
+          name: 'First',
+          description: 'First wins',
+          baseUrl: 'http://localhost:1111',
+        },
+        {
+          id: 'dup-model',
+          name: 'Second',
+          description: 'Dropped',
+          baseUrl: 'http://localhost:2222',
+        },
+      ]),
+      [
+        {
+          id: 'dup-model',
+          name: 'First',
+          description: 'First wins',
+          baseUrl: 'http://localhost:1111',
+        },
+      ],
+    );
+  });
+});
+
+describe('getLocalModels', () => {
+  it('returns only entries with a configured baseUrl', () => {
+    assert.deepEqual(
+      getLocalModels([
+        {
+          id: 'with-url',
+          name: 'With URL',
+          description: 'Has baseUrl',
+          baseUrl: 'http://localhost:1234',
+        },
+        {
+          id: 'without-url',
+          name: 'Without URL',
+          description: 'Missing baseUrl',
+        },
+      ]),
+      [
+        {
+          id: 'with-url',
+          name: 'With URL',
+          description: 'Has baseUrl',
+          baseUrl: 'http://localhost:1234',
+        },
+      ],
+    );
+  });
+
+  it('excludes entries with blank or whitespace-only baseUrl', () => {
+    assert.deepEqual(
+      getLocalModels([
+        {
+          id: 'blank-url',
+          name: 'Blank',
+          description: 'Blank baseUrl',
+          baseUrl: '   ',
+        },
+        {
+          id: 'valid-url',
+          name: 'Valid',
+          description: 'Valid baseUrl',
+          baseUrl: 'http://localhost:1234',
+        },
+      ]),
+      [
+        {
+          id: 'valid-url',
+          name: 'Valid',
+          description: 'Valid baseUrl',
+          baseUrl: 'http://localhost:1234',
+        },
+      ],
+    );
+  });
+
+  it('keeps models with and without apiKey when baseUrl exists', () => {
+    assert.deepEqual(
+      getLocalModels([
+        {
+          id: 'with-api-key',
+          name: 'With API key',
+          description: 'Remote OpenAI-compatible',
+          baseUrl: 'https://api.example.com',
+          apiKey: 'EXAMPLE_API_KEY',
+        },
+        {
+          id: 'without-api-key',
+          name: 'Without API key',
+          description: 'Local server with no auth',
+          baseUrl: 'http://localhost:1234',
+        },
+      ]),
+      [
+        {
+          id: 'with-api-key',
+          name: 'With API key',
+          description: 'Remote OpenAI-compatible',
+          baseUrl: 'https://api.example.com',
+          apiKey: 'EXAMPLE_API_KEY',
+        },
+        {
+          id: 'without-api-key',
+          name: 'Without API key',
+          description: 'Local server with no auth',
+          baseUrl: 'http://localhost:1234',
+        },
+      ],
+    );
+  });
+});
+
+describe('isLocalModelId', () => {
+  it('detects catalog ids', () => {
+    assert.equal(isLocalModelId('qwen2.5-coder-7b', sampleLocalModels), true);
+    assert.equal(isLocalModelId('openai/gpt-5.5', sampleLocalModels), false);
+  });
+});
+
+describe('getAuxiliaryLocalModel', () => {
+  it('returns the model marked useForAux', () => {
+    const models = parseLocalModelsJson([
+      { id: 'local-chat', name: 'Chat', description: 'Chat model' },
+      {
+        id: 'local-aux',
+        name: 'Aux',
+        description: 'Aux model',
+        useForAux: true,
+      },
+    ]);
+    assert.equal(getAuxiliaryLocalModel(models)?.id, 'local-aux');
+  });
+});
+
+describe('localModelToPickerConfig', () => {
+  it('maps catalog entries to picker config', () => {
+    assert.deepEqual(localModelToPickerConfig(sampleLocalModels[0]!), {
+      id: 'qwen2.5-coder-7b',
+      name: 'Qwen',
+      description: 'Local',
+      provider: 'Local',
+      supportsTools: true,
+      supportsThinking: false,
+      supportsVision: false,
+      disabled: false,
+    });
+  });
+
+  it('does not expose apiKey, baseUrl, or supportsForcedToolChoice', () => {
+    const picker = localModelToPickerConfig({
+      id: 'secure-model',
+      name: 'Secure',
+      description: 'Server-only key',
+      baseUrl: 'https://example.local/v1',
+      apiKey: 'DO_NOT_EXPOSE',
+      supportsForcedToolChoice: false,
+    });
+
+    assert.equal('apiKey' in picker, false);
+    assert.equal('baseUrl' in picker, false);
+    assert.equal('supportsForcedToolChoice' in picker, false);
+  });
+});
+
+describe('supportsForcedToolChoice catalog field', () => {
+  it('parses supportsForcedToolChoice independently of supportsThinking', () => {
+    assert.deepEqual(
+      parseLocalModelsJson([
+        {
+          id: 'no-forced-choice',
+          name: 'No Forced',
+          description: 'Rejects forced tool_choice without thinking',
+          baseUrl: 'http://localhost:1234',
+          supportsTools: true,
+          supportsThinking: false,
+          supportsForcedToolChoice: false,
+        },
+      ]),
+      [
+        {
+          id: 'no-forced-choice',
+          name: 'No Forced',
+          description: 'Rejects forced tool_choice without thinking',
+          baseUrl: 'http://localhost:1234',
+          supportsTools: true,
+          supportsThinking: false,
+          supportsForcedToolChoice: false,
+        },
+      ],
+    );
+  });
+});
+
+describe('normalizeOpenRouterBaseUrl', () => {
+  it('returns undefined when unset', () => {
+    assert.equal(normalizeOpenRouterBaseUrl(''), undefined);
+    assert.equal(normalizeOpenRouterBaseUrl('   '), undefined);
+  });
+
+  it('appends /v1 when missing', () => {
+    assert.equal(
+      normalizeOpenRouterBaseUrl('http://localhost:1234'),
+      'http://localhost:1234/v1',
+    );
+  });
+
+  it('preserves an existing /v1 suffix', () => {
+    assert.equal(
+      normalizeOpenRouterBaseUrl('http://localhost:1234/v1/'),
+      'http://localhost:1234/v1',
+    );
+  });
+});
+
+describe('providerFor', () => {
+  it('keeps cloud providers on their direct routes', () => {
+    assert.equal(
+      providerFor('anthropic/claude-opus-4.8', sampleLocalModels),
+      'anthropic',
+    );
+    assert.equal(
+      providerFor('google/gemini-3.1-pro-preview', sampleLocalModels),
+      'google',
+    );
+    assert.equal(
+      providerFor('openai/gpt-5.5', sampleLocalModels),
+      'openrouter',
+    );
+  });
+
+  it('routes catalog ids with a baseUrl locally', () => {
+    assert.equal(providerFor('qwen2.5-coder-7b', sampleLocalModels), 'local');
+  });
+
+  it('falls back to openrouter for catalog ids without a baseUrl', () => {
+    const noUrlModels = parseLocalModelsJson([
+      { id: 'qwen2.5-coder-7b', name: 'Qwen', description: 'Local' },
+    ]);
+    assert.equal(providerFor('qwen2.5-coder-7b', noUrlModels), 'openrouter');
+  });
+});
+
+describe('isMissingLocalBaseUrl', () => {
+  it('is true when catalog lists the model but has no baseUrl', () => {
+    const noUrlModels = parseLocalModelsJson([
+      { id: 'qwen2.5-coder-7b', name: 'Qwen', description: 'Local' },
+    ]);
+    assert.equal(isMissingLocalBaseUrl('qwen2.5-coder-7b', noUrlModels), true);
+    assert.equal(isMissingLocalBaseUrl('openai/gpt-5.5', noUrlModels), false);
+  });
+
+  it('is false when the catalog model has a baseUrl', () => {
+    assert.equal(
+      isMissingLocalBaseUrl('qwen2.5-coder-7b', sampleLocalModels),
+      false,
+    );
+  });
+});
diff --git a/shared/localModels.ts b/shared/localModels.ts
new file mode 100644
index 00000000..60332326
--- /dev/null
+++ b/shared/localModels.ts
@@ -0,0 +1,119 @@
+/**
+ * Pure routing logic for the local OpenAI-compatible model catalog
+ * (`local-models.json`). No env or disk I/O — `localChatConfig.ts` wires
+ * server env + file reads; the client fetches picker entries from `/api/local-models`.
+ */
+import { z } from 'zod';
+
+const modelCoreSchema = z.object({
+  id: z.string().min(1),
+  name: z.string().min(1),
+  description: z.string().min(1),
+  supportsTools: z.boolean().optional(),
+  supportsThinking: z.boolean().optional(),
+  supportsVision: z.boolean().optional(),
+});
+
+// Shape of each entry in `local-models.json`.
+export const localModelSchema = modelCoreSchema.extend({
+  baseUrl: z.string().optional(),
+  apiKey: z.string().optional(),
+  useForAux: z.boolean().optional(),
+  supportsForcedToolChoice: z.boolean().optional(),
+});
+
+export type LocalModelConfig = z.infer<typeof localModelSchema>;
+
+export const pickerModelSchema = modelCoreSchema.extend({
+  provider: z.string().optional(),
+  disabled: z.boolean().optional(),
+});
+
+// UI model-picker shape; cloud entries in `utils.ts` use the same fields.
+export type PickerModelConfig = z.infer<typeof pickerModelSchema>;
+
+export type ChatProvider = 'anthropic' | 'google' | 'openrouter' | 'local';
+
+// Parse and validate a raw `local-models.json` payload; skip invalid entries.
+// skip second occurence of local model with same model id
+export function parseLocalModelsJson(raw: unknown): LocalModelConfig[] {
+  if (!Array.isArray(raw)) return [];
+  const seenIds = new Set<string>();
+  return raw.flatMap((entry) => {
+    const parsed = localModelSchema.safeParse(entry);
+    if (!parsed.success || seenIds.has(parsed.data.id)) return [];
+    seenIds.add(parsed.data.id);
+    return [parsed.data];
+  });
+}
+
+export function isLocalModelId(
+  id: string,
+  localModels: LocalModelConfig[],
+): boolean {
+  return localModels.some((model) => model.id === id);
+}
+
+// One catalog entry may be marked `useForAux` for lightweight tasks (conversation
+// titles, suggestion pills) when `ANTHROPIC_API_KEY` is unset.
+export function getAuxiliaryLocalModel(
+  localModels: LocalModelConfig[],
+): LocalModelConfig | undefined {
+  return localModels.find((model) => model.useForAux === true);
+}
+
+// Vercel AI SDK OpenRouter provider expects baseURL to include `/v1` (its default
+// is `https://openrouter.ai/api/v1`). LM Studio and similar servers often
+// advertise the bare host; append `/v1` when missing.
+export function normalizeOpenRouterBaseUrl(raw: string): string | undefined {
+  const trimmed = raw.trim();
+  if (!trimmed) return undefined;
+  const base = trimmed.replace(/\/+$/, '');
+  return base.endsWith('/v1') ? base : `${base}/v1`;
+}
+
+// Cloud ids with a provider prefix keep their direct routes; catalog ids with a
+// configured baseUrl route to their own OpenAI-compatible server; everything
+// else goes through OpenRouter.
+export function providerFor(
+  modelId: string,
+  localModels: LocalModelConfig[],
+): ChatProvider {
+  if (modelId.startsWith('anthropic/')) return 'anthropic';
+  if (modelId.startsWith('google/')) return 'google';
+  const localModel = localModels.find((m) => m.id === modelId);
+  if (localModel?.baseUrl?.trim()) return 'local';
+  return 'openrouter';
+}
+
+// Catalog ids without a baseUrl cannot be routed — callers reject those
+// requests with a clear error instead of falling through to OpenRouter.
+export function isMissingLocalBaseUrl(
+  modelId: string,
+  catalogOnDisk: LocalModelConfig[],
+): boolean {
+  const model = catalogOnDisk.find((m) => m.id === modelId);
+  return !!model && !model.baseUrl?.trim();
+}
+
+// Parametric CAD requires tool calling — grey out catalog entries that explicitly
+// opt out via `supportsTools: false`.
+export function localModelToPickerConfig(
+  model: LocalModelConfig,
+): PickerModelConfig {
+  return {
+    id: model.id,
+    name: model.name,
+    description: model.description,
+    provider: 'Local',
+    supportsTools: model.supportsTools ?? true,
+    supportsThinking: model.supportsThinking ?? false,
+    supportsVision: model.supportsVision ?? false,
+    disabled: model.supportsTools === false,
+  };
+}
+
+// Returns parsed, valid catalog entries that have a baseUrl configured.
+export function getLocalModels(rawJson: unknown): LocalModelConfig[] {
+  return parseLocalModelsJson(rawJson).filter((m) => !!m.baseUrl?.trim());
+}
diff --git a/src/components/ModelSelector.tsx b/src/components/ModelSelector.tsx
index fd5cb3ae..96bb0864 100644
--- a/src/components/ModelSelector.tsx
+++ b/src/components/ModelSelector.tsx
@@ -7,6 +7,7 @@ import {
   DropdownMenuTrigger,
 } from '@/components/ui/dropdown-menu';
 import { Button } from '@/components/ui/button';
+import { ScrollArea } from '@/components/ui/scroll-area';
 import { cn } from '@/lib/utils';
 import { Model } from '@shared/types';
 import { ModelConfig } from '../types/misc.ts';
@@ -174,7 +175,7 @@ export function ModelSelector({
         </Button>
       </DropdownMenuTrigger>
       <DropdownMenuContent
-        className="flex w-64 flex-col gap-1 rounded-lg bg-adam-neutral-700 p-1"
+        className="w-64 rounded-lg bg-adam-neutral-700 p-1"
         align="end"
         // Radix fires this just before it tries to refocus the trigger.
         // We intercept the event: if the menu was opened with a pointer we
@@ -189,47 +190,52 @@ export function ModelSelector({
           }
         }}
       >
-        {models.map((model) => (
-          <DropdownMenuItem
-            key={model.id}
-            className={cn(
-              'cursor-pointer rounded-md bg-adam-neutral-700 px-4 transition-colors duration-150 focus:bg-adam-bg-secondary-dark',
-              currentType === 'creative' ? 'py-3' : 'py-2',
-              selectedModel === model.id && 'bg-adam-neutral-800',
-              !!model.disabled && 'cursor-not-allowed opacity-50',
-            )}
-            onClick={(event) => {
-              onModelChange(model.id);
-              setIsDropdownOpen(false);
-              event.stopPropagation();
-            }}
-            disabled={!!model.disabled}
-          >
-            <div className="flex-1">
-              <span
+        <ScrollArea className="[&_[data-radix-scroll-area-viewport]]:max-h-[360px] [&_[data-radix-scroll-area-viewport]]:overflow-x-hidden">
+          <div className="flex flex-col gap-1">
+            {models.map((model) => (
+              <DropdownMenuItem
+                key={model.id}
                 className={cn(
-                  'text-sm font-medium',
-                  focused ? 'text-white' : 'text-adam-text-primary',
+                  'cursor-pointer rounded-md bg-adam-neutral-700 px-4 py-3 transition-colors duration-150 focus:bg-adam-bg-secondary-dark',
+                  selectedModel === model.id && 'bg-adam-neutral-800',
+                  !!model.disabled && 'cursor-not-allowed opacity-50',
                 )}
+                onClick={(event) => {
+                  onModelChange(model.id);
+                  setIsDropdownOpen(false);
+                  event.stopPropagation();
+                }}
+                disabled={!!model.disabled}
               >
-                {model.name}
-              </span>
-              {/* Creative modes (Draft/Textureless/Max Quality) rely on the
-                  description to explain the speed/quality/texture tradeoff;
-                  parametric models show names only to keep the list compact. */}
-              {currentType === 'creative' && model.description && (
-                <p
-                  className={cn(
-                    'mt-0.5 text-xs',
-                    focused ? 'text-white' : 'text-gray-400',
+                <div className="flex-1">
+                  <div className="flex items-center">
+                    <span
+                      className={cn(
+                        'font-medium',
+                        focused ? 'text-white' : 'text-adam-text-primary',
+                      )}
+                    >
+                      {model.name}
+                    </span>
+                  </div>
+                  {/* Creative modes (Draft/Textureless/Max Quality) rely on the
+                    description to explain the speed/quality/texture tradeoff;
+                    parametric models show names only to keep the list compact. */}
+                  {currentType === 'creative' && model.description && (
+                    <span
+                      className={cn(
+                        'mt-0.5 text-xs',
+                        focused ? 'text-white' : 'text-gray-400',
+                      )}
+                    >
+                      {model.description}
+                    </span>
                   )}
-                >
-                  {model.description}
-                </p>
-              )}
-            </div>
-          </DropdownMenuItem>
-        ))}
+                </div>
+              </DropdownMenuItem>
+            ))}
+          </div>
+        </ScrollArea>
       </DropdownMenuContent>
     </DropdownMenu>
   );
diff --git a/src/components/TextAreaChat.tsx b/src/components/TextAreaChat.tsx
index 23a435d5..a950eb59 100644
--- a/src/components/TextAreaChat.tsx
+++ b/src/components/TextAreaChat.tsx
@@ -20,9 +20,9 @@ import {
 import {
   cn,
   CREATIVE_MODELS,
-  PARAMETRIC_MODELS,
   parametricModelSupportsVision,
 } from '@/lib/utils';
+import { useParametricModels } from '@/hooks/useLocalModels.ts';
 import { CreativeModel, MeshFileType, Model } from '@shared/types';
 import type { AppUIMessage } from '@shared/chatAi';
 import { imageFilePartUrl } from '@shared/imageRefs';
@@ -652,12 +652,13 @@ function TextAreaChat({
     },
   };
 
+  const parametricModels = useParametricModels();
   const memoizedModels = useMemo(() => {
     if (type === 'creative') {
       return CREATIVE_MODELS;
     }
-    return PARAMETRIC_MODELS;
-  }, [type]);
+    return parametricModels;
+  }, [type, parametricModels]);
 
   // ------------------------------------------------------------
   // Placeholder – Typed-out Animation
@@ -1594,7 +1595,7 @@ function TextAreaChat({
         <div className="flex items-center justify-between border-t border-[#2a2a2a] p-3">
           <div className="flex items-center gap-1">
             {(type !== 'parametric' ||
-              parametricModelSupportsVision(model)) && (
+              parametricModelSupportsVision(model, memoizedModels)) && (
               <div
                 className={cn(
                   'transition-all duration-300 ease-out',
diff --git a/src/components/chat/ChatSession.tsx b/src/components/chat/ChatSession.tsx
index f19d9eae..611819cd 100644
--- a/src/components/chat/ChatSession.tsx
+++ b/src/components/chat/ChatSession.tsx
@@ -29,6 +29,7 @@ import {
 } from 'ai';
 import Tree from '@shared/Tree';
 import { isParametricArtifact } from '@shared/parametricParts';
+import { inspectionPreviewStoragePath } from '@shared/imageRefs';
 import type {
   Conversation,
   Message,
@@ -470,7 +471,11 @@ export function ChatSession({
             const inspectionBlob = await fetch(inspectionDataUrl).then(
               (response) => response.blob(),
             );
-            const inspectionPath = `${user.id}/${conversation.id}/inspection-preview-${toolCall.toolCallId}`;
+            const inspectionPath = inspectionPreviewStoragePath(
+              user.id,
+              conversation.id,
+              toolCall.toolCallId,
+            );
             const { error: inspectionUploadError } = await supabase.storage
               .from('images')
               .upload(inspectionPath, inspectionBlob, {
diff --git a/src/components/chat/MessageBubble.tsx b/src/components/chat/MessageBubble.tsx
index d1a13867..21c64941 100644
--- a/src/components/chat/MessageBubble.tsx
+++ b/src/components/chat/MessageBubble.tsx
@@ -8,7 +8,8 @@ import {
   DropdownMenuItem,
   DropdownMenuTrigger,
 } from '@/components/ui/dropdown-menu';
-import { CREATIVE_MODELS, PARAMETRIC_MODELS } from '@/lib/utils';
+import { CREATIVE_MODELS } from '@/lib/utils';
+import { useParametricModels } from '@/hooks/useLocalModels';
 import { Avatar, AvatarImage } from '@/components/ui/avatar';
 import { Button } from '@/components/ui/button';
 import { ScrollArea } from '@/components/ui/scroll-area';
@@ -430,8 +431,9 @@ function AssistantBubble({
   onRestore,
 }: MessageBubbleProps) {
   const { conversation } = useConversation();
+  const parametricModels = useParametricModels();
   const modelOptions =
-    conversation.type === 'creative' ? CREATIVE_MODELS : PARAMETRIC_MODELS;
+    conversation.type === 'creative' ? CREATIVE_MODELS : parametricModels;
   const [expandedTools, setExpandedTools] = useState<Set<number>>(new Set());
   const lastParametricBuildIndex = useMemo(() => {
     if (conversation.type !== 'parametric') return -1;
@@ -836,18 +838,20 @@ function RetryModelDropdown({
         align="end"
         className="w-48 rounded-lg border border-adam-neutral-700 bg-adam-neutral-800 p-1"
       >
-        {modelOptions.map((option) => (
-          <DropdownMenuItem
-            key={option.id}
-            className="cursor-pointer rounded-md bg-adam-neutral-800 px-2 py-1.5 text-xs text-adam-text-primary hover:bg-adam-neutral-700 focus:bg-adam-bg-secondary-dark"
-            onClick={() => {
-              onRetry(option.id);
-              setIsOpen(false);
-            }}
-          >
-            {option.name}
-          </DropdownMenuItem>
-        ))}
+        <ScrollArea className="[&_[data-radix-scroll-area-viewport]]:max-h-[140px] [&_[data-radix-scroll-area-viewport]]:overflow-x-hidden">
+          {modelOptions.map((option) => (
+            <DropdownMenuItem
+              key={option.id}
+              className="cursor-pointer rounded-md bg-adam-neutral-800 px-2 py-1.5 text-xs text-adam-text-primary hover:bg-adam-neutral-700 focus:bg-adam-bg-secondary-dark"
+              onClick={() => {
+                onRetry(option.id);
+                setIsOpen(false);
+              }}
+            >
+              {option.name}
+            </DropdownMenuItem>
+          ))}
+        </ScrollArea>
       </DropdownMenuContent>
     </DropdownMenu>
   );
diff --git a/src/hooks/useLocalModels.ts b/src/hooks/useLocalModels.ts
new file mode 100644
index 00000000..8bed6eeb
--- /dev/null
+++ b/src/hooks/useLocalModels.ts
@@ -0,0 +1,28 @@
+import { useMemo } from 'react';
+import { useQuery } from '@tanstack/react-query';
+import { z } from 'zod';
+import { isLocalModelId, pickerModelSchema } from '@shared/localModels';
+import { PARAMETRIC_MODELS } from '@/lib/utils';
+import { apiJson } from '@/services/api';
+import type { ModelConfig } from '@/types/misc';
+
+const localModelsSchema = z.array(pickerModelSchema);
+
+function useLocalModelsQuery() {
+  return useQuery({
+    queryKey: ['local-models'],
+    queryFn: () => apiJson('local-models', {}, localModelsSchema),
+    staleTime: 5 * 60 * 1000,
+  });
+}
+
+export function useIsLocalModel(modelId: string): boolean {
+  const { data: localModels = [] } = useLocalModelsQuery();
+  return isLocalModelId(modelId, localModels);
+}
+
+export function useParametricModels(): ModelConfig[] {
+  const { data: localModels = [] } = useLocalModelsQuery();
+
+  return useMemo(() => [...PARAMETRIC_MODELS, ...localModels], [localModels]);
+}
diff --git a/src/lib/utils.ts b/src/lib/utils.ts
index 3f3e0fb8..35cb326f 100644
--- a/src/lib/utils.ts
+++ b/src/lib/utils.ts
@@ -235,7 +235,7 @@ export function getInitials(fullName: string | null) {
   return 'U';
 }
 
-export const PARAMETRIC_MODELS: ModelConfig[] = [
+const CLOUD_PARAMETRIC_MODELS: ModelConfig[] = [
   {
     id: 'google/gemini-3.1-pro-preview',
     name: 'Gemini 3.1 Pro',
@@ -292,6 +292,8 @@ export const PARAMETRIC_MODELS: ModelConfig[] = [
   },
 ];
 
+export const PARAMETRIC_MODELS: ModelConfig[] = [...CLOUD_PARAMETRIC_MODELS];
+
 export const CREATIVE_MODELS: ModelConfig[] = [
   {
     id: 'ultra',
@@ -316,7 +318,10 @@ export const CREATIVE_MODELS: ModelConfig[] = [
 // Whether the selected parametric model can accept image / STL-render inputs.
 // Unknown ids (e.g. historical messages tagged with a removed model) fall back
 // to `true` so older saved rows still render normally.
-export function parametricModelSupportsVision(modelId: string): boolean {
-  const cfg = PARAMETRIC_MODELS.find((m) => m.id === modelId);
+export function parametricModelSupportsVision(
+  modelId: string,
+  models: ModelConfig[] = PARAMETRIC_MODELS,
+): boolean {
+  const cfg = models.find((m) => m.id === modelId);
   return cfg?.supportsVision !== false;
 }
diff --git a/src/routeTree.gen.ts b/src/routeTree.gen.ts
index ad98721d..0d725953 100644
--- a/src/routeTree.gen.ts
+++ b/src/routeTree.gen.ts
@@ -23,6 +23,7 @@ import { Route as ApiTitleGeneratorRouteImport } from './routes/api/title-genera
 import { Route as ApiPromptGeneratorRouteImport } from './routes/api/prompt-generator'
 import { Route as ApiParametricChatRouteImport } from './routes/api/parametric-chat'
 import { Route as ApiMeshRouteImport } from './routes/api/mesh'
+import { Route as ApiLocalModelsRouteImport } from './routes/api/local-models'
 import { Route as ApiFalWebhookRouteImport } from './routes/api/fal-webhook'
 import { Route as ApiDeleteUserRouteImport } from './routes/api/delete-user'
 import { Route as ApiCreativeChatRouteImport } from './routes/api/creative-chat'
@@ -108,6 +109,11 @@ const ApiMeshRoute = ApiMeshRouteImport.update({
   path: '/api/mesh',
   getParentRoute: () => rootRouteImport,
 } as any)
+const ApiLocalModelsRoute = ApiLocalModelsRouteImport.update({
+  id: '/api/local-models',
+  path: '/api/local-models',
+  getParentRoute: () => rootRouteImport,
+} as any)
 const ApiFalWebhookRoute = ApiFalWebhookRouteImport.update({
   id: '/api/fal-webhook',
   path: '/api/fal-webhook',
@@ -201,6 +207,7 @@ export interface FileRoutesByFullPath {
   '/api/creative-chat': typeof ApiCreativeChatRoute
   '/api/delete-user': typeof ApiDeleteUserRoute
   '/api/fal-webhook': typeof ApiFalWebhookRoute
+  '/api/local-models': typeof ApiLocalModelsRoute
   '/api/mesh': typeof ApiMeshRoute
   '/api/parametric-chat': typeof ApiParametricChatRoute
   '/api/prompt-generator': typeof ApiPromptGeneratorRoute
@@ -230,6 +237,7 @@ export interface FileRoutesByTo {
   '/api/creative-chat': typeof ApiCreativeChatRoute
   '/api/delete-user': typeof ApiDeleteUserRoute
   '/api/fal-webhook': typeof ApiFalWebhookRoute
+  '/api/local-models': typeof ApiLocalModelsRoute
   '/api/mesh': typeof ApiMeshRoute
   '/api/parametric-chat': typeof ApiParametricChatRoute
   '/api/prompt-generator': typeof ApiPromptGeneratorRoute
@@ -261,6 +269,7 @@ export interface FileRoutesById {
   '/api/creative-chat': typeof ApiCreativeChatRoute
   '/api/delete-user': typeof ApiDeleteUserRoute
   '/api/fal-webhook': typeof ApiFalWebhookRoute
+  '/api/local-models': typeof ApiLocalModelsRoute
   '/api/mesh': typeof ApiMeshRoute
   '/api/parametric-chat': typeof ApiParametricChatRoute
   '/api/prompt-generator': typeof ApiPromptGeneratorRoute
@@ -293,6 +302,7 @@ export interface FileRouteTypes {
     | '/api/creative-chat'
     | '/api/delete-user'
     | '/api/fal-webhook'
+    | '/api/local-models'
     | '/api/mesh'
     | '/api/parametric-chat'
     | '/api/prompt-generator'
@@ -322,6 +332,7 @@ export interface FileRouteTypes {
     | '/api/creative-chat'
     | '/api/delete-user'
     | '/api/fal-webhook'
+    | '/api/local-models'
     | '/api/mesh'
     | '/api/parametric-chat'
     | '/api/prompt-generator'
@@ -352,6 +363,7 @@ export interface FileRouteTypes {
     | '/api/creative-chat'
     | '/api/delete-user'
     | '/api/fal-webhook'
+    | '/api/local-models'
     | '/api/mesh'
     | '/api/parametric-chat'
     | '/api/prompt-generator'
@@ -382,6 +394,7 @@ export interface RootRouteChildren {
   ApiCreativeChatRoute: typeof ApiCreativeChatRoute
   ApiDeleteUserRoute: typeof ApiDeleteUserRoute
   ApiFalWebhookRoute: typeof ApiFalWebhookRoute
+  ApiLocalModelsRoute: typeof ApiLocalModelsRoute
   ApiMeshRoute: typeof ApiMeshRoute
   ApiParametricChatRoute: typeof ApiParametricChatRoute
   ApiPromptGeneratorRoute: typeof ApiPromptGeneratorRoute
@@ -489,6 +502,13 @@ declare module '@tanstack/react-router' {
       preLoaderRoute: typeof ApiMeshRouteImport
       parentRoute: typeof rootRouteImport
     }
+    '/api/local-models': {
+      id: '/api/local-models'
+      path: '/api/local-models'
+      fullPath: '/api/local-models'
+      preLoaderRoute: typeof ApiLocalModelsRouteImport
+      parentRoute: typeof rootRouteImport
+    }
     '/api/fal-webhook': {
       id: '/api/fal-webhook'
       path: '/api/fal-webhook'
@@ -649,6 +669,7 @@ const rootRouteChildren: RootRouteChildren = {
   ApiCreativeChatRoute: ApiCreativeChatRoute,
   ApiDeleteUserRoute: ApiDeleteUserRoute,
   ApiFalWebhookRoute: ApiFalWebhookRoute,
+  ApiLocalModelsRoute: ApiLocalModelsRoute,
   ApiMeshRoute: ApiMeshRoute,
   ApiParametricChatRoute: ApiParametricChatRoute,
   ApiPromptGeneratorRoute: ApiPromptGeneratorRoute,
diff --git a/src/routes/api/local-models.ts b/src/routes/api/local-models.ts
new file mode 100644
index 00000000..ef27024f
--- /dev/null
+++ b/src/routes/api/local-models.ts
@@ -0,0 +1,16 @@
+import { createFileRoute } from '@tanstack/react-router';
+import { localModelToPickerConfig } from '@shared/localModels';
+import { getLocalChatState } from '@/server/localChatConfig';
+import { json, preflight } from '@/server/api';
+
+export const Route = createFileRoute('/api/local-models')({
+  server: {
+    handlers: {
+      OPTIONS: preflight,
+      GET: async () => {
+        const { activeCatalog } = getLocalChatState();
+        return json(activeCatalog.map(localModelToPickerConfig));
+      },
+    },
+  },
+});
diff --git a/src/server/aiChat.ts b/src/server/aiChat.ts
index 3066bd1c..36eba748 100644
--- a/src/server/aiChat.ts
+++ b/src/server/aiChat.ts
@@ -2,8 +2,16 @@ import { createAnthropic } from '@ai-sdk/anthropic';
 import { createGoogleGenerativeAI } from '@ai-sdk/google';
 import { createOpenRouter } from '@openrouter/ai-sdk-provider';
 import { chatTools, type AppUIMessage, type AppTools } from '@shared/chatAi';
+import {
+  getAuxiliaryLocalModel,
+  normalizeOpenRouterBaseUrl,
+} from '@shared/localModels';
 import { cleanAssistantText, getParametricText } from '@shared/parametricParts';
-import { imageIdFromFilename, imageStoragePath } from '@shared/imageRefs';
+import {
+  imageIdFromFilename,
+  imageStoragePath,
+  inspectionPreviewStoragePath,
+} from '@shared/imageRefs';
 import { normalizeConversationSuggestions } from '@shared/suggestions';
 import type { Conversation, Message, MeshFileType, Model } from '@shared/types';
 import {
@@ -34,7 +42,17 @@ import {
   resolveDanglingToolParts,
 } from './chatToolPersistence';
 import { handleMeshRequest } from './mesh';
+import {
+  getActiveLocalModel,
+  getLocalChatState,
+  isActiveLocalModel,
+  isMissingLocalBaseUrl,
+  localApiKeyForModel,
+  providerFor,
+  type ChatProvider,
+} from './localChatConfig';
 import { getAnonSupabaseClient } from './supabaseClient';
+import { findLatestBuildToolCallId } from './messageUtils';
 
 /**
  * USD list price per **million** tokens, keyed by the same model IDs the
@@ -309,16 +327,9 @@ function jsonResponse(body: unknown, status: number) {
 const THINKING_BUDGET_TOKENS = 9000;
 const PARAMETRIC_MAX_OUTPUT_TOKENS = 64000;
 
-type ChatProvider = 'anthropic' | 'google' | 'openrouter';
-
-function providerFor(modelId: string): ChatProvider {
-  if (modelId.startsWith('anthropic/')) return 'anthropic';
-  if (modelId.startsWith('google/')) return 'google';
-  return 'openrouter';
-}
-
 type AnthropicProvider = ReturnType<typeof createAnthropic>;
 type GoogleProvider = ReturnType<typeof createGoogleGenerativeAI>;
+type OpenRouterProvider = ReturnType<typeof createOpenRouter>;
 
 // The Vercel AI SDK's Anthropic provider expects ANTHROPIC_BASE_URL to already
 // include the "/v1" path segment (its built-in default is
@@ -337,13 +348,15 @@ function normalizedAnthropicBaseURL(): string | undefined {
 type ChatProviders = {
   anthropic: () => AnthropicProvider;
   google: () => GoogleProvider;
-  openrouter: () => ReturnType<typeof createOpenRouter>;
+  openrouter: () => OpenRouterProvider;
+  local: (baseUrl: string, apiKey?: string) => OpenRouterProvider;
 };
 
 function createChatProviders(): ChatProviders {
   let anthropic: AnthropicProvider | undefined;
   let google: GoogleProvider | undefined;
-  let openrouter: ReturnType<typeof createOpenRouter> | undefined;
+  let openrouter: OpenRouterProvider | undefined;
+  const localByTarget = new Map<string, OpenRouterProvider>();
   return {
     anthropic: () => {
       if (!anthropic) {
@@ -367,6 +380,20 @@ function createChatProviders(): ChatProviders {
       });
       return openrouter;
     },
+    local: (baseUrl: string, apiKey?: string) => {
+      const resolvedApiKey = apiKey?.trim() || 'local';
+      const providerKey = `${baseUrl}::${resolvedApiKey}`;
+      let provider = localByTarget.get(providerKey);
+      if (!provider) {
+        provider = createOpenRouter({
+          baseURL: baseUrl,
+          apiKey: resolvedApiKey,
+          compatibility: 'compatible',
+        });
+        localByTarget.set(providerKey, provider);
+      }
+      return provider;
+    },
   };
 }
 
@@ -386,8 +413,22 @@ function buildChatModel(
 ): { model: LanguageModel; providerOptions?: ProviderOptions } {
   const hasCappedThinkingBudget =
     thinking && thinkingBudget !== THINKING_BUDGET_TOKENS;
+  const provider = providerFor(modelId);
+
+  if (provider === 'local') {
+    const modelConfig = getActiveLocalModel(modelId);
+    const rawBaseUrl = modelConfig?.baseUrl;
+    if (!rawBaseUrl)
+      throw new Error(`No baseUrl configured for local model ${modelId}`);
+    const baseUrl = normalizeOpenRouterBaseUrl(rawBaseUrl) ?? rawBaseUrl;
+    return {
+      model: providers
+        .local(baseUrl, localApiKeyForModel(modelConfig))
+        .chat(modelId),
+    };
+  }
 
-  if (providerFor(modelId) === 'openrouter') {
+  if (provider === 'openrouter') {
     return {
       model: providers.openrouter().chat(modelId, {
         ...(thinking ? { reasoning: { max_tokens: thinkingBudget } } : {}),
@@ -457,6 +498,14 @@ function bareModelId(modelId: string): string {
   return id.replace(/\./g, '-');
 }
 
+// For local models, check supportsVision.
+// Otherwise, all cloud models support vision.
+function chatModelSupportsVision(modelId: string): boolean {
+  const local = getActiveLocalModel(modelId);
+  if (local) return local.supportsVision === true;
+  return true;
+}
+
 // The Claude 5 generation swaps the opus/sonnet/haiku tiers for code names
 // ("claude-fable-5", "claude-mythos-5", …). Match the `claude-<codename>-5`
 // shape rather than enumerating code names so future Claude 5 variants
@@ -475,13 +524,24 @@ function usesAdaptiveAnthropicThinking(modelId: string) {
   return match ? Number(match[1]) >= 6 : false;
 }
 
-// The reasoning-tier Claude 5 models (Fable, Mythos) reject a forced
-// `tool_choice` outright ("tool_choice forces tool use is not compatible with
-// this model"). Other Claude 5 tiers — notably Sonnet 5 — accept a forced
-// tool_choice on the first-party API, provided thinking is disabled for that
-// step (see the per-step override in the parametric flow).
+// The reasoning-tier Claude 5 models (Fable, Mythos) and some local /
+// OpenAI-compatible servers reject a forced `tool_choice` outright
+// ("tool_choice forces tool use is not compatible with this model"). Other
+// Claude 5 tiers — notably Sonnet 5 — accept a forced tool_choice on the
+// first-party API, provided thinking is disabled for that step (see the
+// per-step override in the parametric flow).
+//
+// For local catalog entries, `supportsForcedToolChoice` is the explicit
+// knob. When unset, fall back to treating `supportsThinking: true` as a
+// rejection signal (common for thinking-capable OpenAI-compatible servers).
 function rejectsForcedToolChoice(modelId: string): boolean {
-  return /^claude-(?:fable|mythos)\b/.test(bareModelId(modelId));
+  if (/^claude-(?:fable|mythos)\b/.test(bareModelId(modelId))) return true;
+  const local = getActiveLocalModel(modelId);
+  if (!local) return false;
+  if (local.supportsForcedToolChoice !== undefined) {
+    return local.supportsForcedToolChoice === false;
+  }
+  return local.supportsThinking === true;
 }
 
 // Whether a model accepts a forced `tool_choice` (type: "tool" / "any").
@@ -490,6 +550,14 @@ function supportsForcedToolChoice(modelId: string): boolean {
 }
 
 function priceFor(modelId: string) {
+  if (isActiveLocalModel(modelId)) {
+    return {
+      input: 0,
+      output: 0,
+      cacheRead: 0,
+      cacheWrite: 0,
+    };
+  }
   const entry = MODEL_PRICES[modelId] ?? FALLBACK_MODEL_PRICE;
   return {
     input: entry.input,
@@ -543,6 +611,7 @@ function billingTokensFromUsage(
   modelId: string,
   usage: LanguageModelUsage,
 ): number {
+  if (isActiveLocalModel(modelId)) return 0;
   const usdCost = usdCostFromUsage(modelId, usage) * billingMultiplier();
   return Math.max(1, Math.ceil(usdCost / USD_PER_BILLING_TOKEN));
 }
@@ -712,16 +781,16 @@ async function loadBranchFromDb({
 }
 
 async function generateConversationTitle({
-  anthropic,
+  auxModel,
   firstMessage,
 }: {
-  anthropic: AnthropicProvider;
+  auxModel: LanguageModel;
   firstMessage: AppUIMessage;
 }) {
   const text = getParametricText(firstMessage.parts) || 'New conversation';
   try {
     const result = await generateText({
-      model: anthropic('claude-haiku-4-5'),
+      model: auxModel,
       system:
         'Generate a short title for a 3D creation conversation. Return only the title.',
       prompt: text,
@@ -744,11 +813,11 @@ async function generateConversationTitle({
  * specific assistant turn.
  */
 async function generateConversationSuggestions({
-  anthropic,
+  auxModel,
   branch,
   conversationType,
 }: {
-  anthropic: AnthropicProvider;
+  auxModel: LanguageModel;
   branch: AppUIMessage[];
   conversationType: 'parametric' | 'creative';
 }): Promise<string[]> {
@@ -766,7 +835,7 @@ async function generateConversationSuggestions({
   const summary = `User request: ${firstUserText.slice(0, 400)}\n\nMost recent assistant reply: ${lastAssistantText.slice(0, 400)}`;
   try {
     const result = await generateText({
-      model: anthropic('claude-haiku-4-5'),
+      model: auxModel,
       system:
         conversationType === 'creative'
           ? 'Given a 3D mesh design conversation, return an array of exactly 2 follow-up prompts the user might want to send next. Each prompt is a concise instruction of 3 words or fewer, not a question. Return exactly 2 items — no more, no fewer.'
@@ -866,6 +935,125 @@ async function sniffImageMediaType(bytes: Uint8Array): Promise<string | null> {
   return sniffed && ACCEPTED_IMAGE_MEDIA_TYPES.has(sniffed) ? sniffed : null;
 }
 
+async function buildHydratedMessages(
+  messages: AppUIMessage[],
+  {
+    supabaseClient,
+    conversation,
+    modelId,
+    provider,
+  }: {
+    supabaseClient: SupabaseAnon;
+    conversation: ConversationAccess;
+    modelId: string;
+    provider: ChatProvider;
+  },
+): Promise<Array<Omit<AppUIMessage, 'id'>>> {
+  const shouldAddInspection =
+    provider === 'local' &&
+    conversation.type === 'parametric' &&
+    chatModelSupportsVision(modelId);
+  const latestBuildToolCallId = shouldAddInspection
+    ? findLatestBuildToolCallId(messages)
+    : null;
+
+  const hydrated = await Promise.all(
+    messages.map(async (message) => ({
+      ...message,
+      parts: (
+        await Promise.all(
+          message.parts.map(async (part) => {
+            if (
+              part.type !== 'file' ||
+              typeof part.mediaType !== 'string' ||
+              !part.mediaType.startsWith('image/') ||
+              part.url.startsWith('data:')
+            ) {
+              return part;
+            }
+            const imageId = imageIdFromFilename(part.filename);
+            if (!imageId) return null;
+            const downloaded = await downloadAsBase64(
+              supabaseClient,
+              'images',
+              imageStoragePath(conversation.user_id, conversation.id, imageId),
+            );
+            if (!downloaded) return null;
+            return {
+              ...part,
+              mediaType: downloaded.mediaType,
+              url: `data:${downloaded.mediaType};base64,${downloaded.base64}`,
+            };
+          }),
+        )
+      ).filter((part): part is NonNullable<typeof part> => part != null),
+    })),
+  );
+
+  if (!latestBuildToolCallId) {
+    return hydrated.map(({ id: _id, ...message }) => message);
+  }
+
+  const result: Array<Omit<AppUIMessage, 'id'>> = [];
+  for (const message of hydrated) {
+    const { id: _id, ...messageWithoutId } = message;
+    result.push(messageWithoutId);
+
+    if (message.role !== 'assistant') continue;
+
+    const hasLatestBuild = message.parts.some(
+      (part) =>
+        part.type === 'tool-build_parametric_model' &&
+        part.state === 'output-available' &&
+        'toolCallId' in part &&
+        part.toolCallId === latestBuildToolCallId,
+    );
+    if (!hasLatestBuild) continue;
+
+    const inspectionPath = inspectionPreviewStoragePath(
+      conversation.user_id,
+      conversation.id,
+      latestBuildToolCallId,
+    );
+    const downloaded = await downloadAsBase64(
+      supabaseClient,
+      'images',
+      inspectionPath,
+    );
+    if (!downloaded) {
+      logError(new Error('Failed to hydrate inspection preview image'), {
+        functionName: 'ai-chat',
+        statusCode: 500,
+        userId: conversation.user_id,
+        conversationId: conversation.id,
+        additionalContext: {
+          operation: 'inspection_hydration_download',
+          toolCallId: latestBuildToolCallId,
+          storagePath: inspectionPath,
+        },
+      });
+      continue;
+    }
+
+    result.push({
+      role: 'user',
+      parts: [
+        {
+          type: 'text',
+          text: 'Multi-view inspection render from the tool result above.',
+        },
+        {
+          type: 'file',
+          mediaType: downloaded.mediaType,
+          url: `data:${downloaded.mediaType};base64,${downloaded.base64}`,
+        },
+      ],
+    });
+  }
+
+  return result;
+}
+
 async function downloadAsBase64(
   supabaseClient: SupabaseAnon,
   bucket: string,
@@ -892,12 +1080,25 @@ async function downloadAsBase64(
   return { base64: btoa(binary), mediaType };
 }
 
+async function storageObjectExists(
+  supabaseClient: SupabaseAnon,
+  bucket: string,
+  path: string,
+): Promise<boolean> {
+  const { data, error } = await supabaseClient.storage
+    .from(bucket)
+    .createSignedUrl(path, 60);
+  return !error && !!data?.signedUrl;
+}
+
 function parametricTools({
-  previewPathForToolCall,
   supabaseClient,
+  conversation,
+  modelId,
 }: {
-  previewPathForToolCall: (toolCallId: string) => string;
   supabaseClient: SupabaseAnon;
+  conversation: ConversationAccess;
+  modelId: string;
 }) {
   return {
     build_parametric_model: {
@@ -909,42 +1110,94 @@ function parametricTools({
         toolCallId: string;
         output: AppTools['build_parametric_model']['output'];
       }) {
-        // The client uploads a multi-view render of the compiled SCAD to a path
-        // derived from toolCallId BEFORE sending the tool result (see
-        // ChatSession's `onToolCall`). If for any reason the upload
-        // didn't land, `downloadAsBase64` returns null and we fall back
-        // to text-only — never block the loop on a missing inspection sheet.
-        const downloaded = await downloadAsBase64(
-          supabaseClient,
-          'images',
-          previewPathForToolCall(toolCallId),
-        );
         const views =
           output.inspection?.views.join(', ') ??
           'ISO, FRONT, BACK, LEFT, RIGHT, TOP, BOTTOM';
-        const text = `${output.message}\nRendered inspection views: ${views}.\nMulti-view inspection image attached: ${downloaded ? 'yes' : 'no'}.`;
 
-        if (downloaded) {
-          return {
-            type: 'content' as const,
-            value: [
-              { type: 'text' as const, text },
-              {
-                type: 'image-data' as const,
-                data: downloaded.base64,
-                mediaType: downloaded.mediaType,
-              },
-            ],
-          };
-        }
+        if (
+          providerFor(modelId) !== 'local' &&
+          chatModelSupportsVision(modelId)
+        ) {
+          // The client uploads a multi-view render of the compiled SCAD to a path
+          // derived from toolCallId BEFORE sending the tool result (see
+          // ChatSession's `onToolCall`). If for any reason the upload
+          // didn't land, `downloadAsBase64` returns null and we fall back
+          // to text-only — never block the loop on a missing inspection sheet.
+          const downloaded = await downloadAsBase64(
+            supabaseClient,
+            'images',
+            inspectionPreviewStoragePath(
+              conversation.user_id,
+              conversation.id,
+              toolCallId,
+            ),
+          );
+          const text = `${output.message}\nRendered inspection views: ${views}.\nMulti-view inspection image attached: ${downloaded ? 'yes' : 'no'}.`;
+
+          if (downloaded) {
+            return {
+              type: 'content' as const,
+              value: [
+                { type: 'text' as const, text },
+                {
+                  type: 'image-data' as const,
+                  data: downloaded.base64,
+                  mediaType: downloaded.mediaType,
+                },
+              ],
+            };
+          }
 
-        return { type: 'text' as const, value: text };
+          return { type: 'text' as const, value: text };
+        } else {
+          // For models that do not support multimodal tools result messages,
+          // the rendered inspection views are appended as user messages in buildHydratedMessages
+
+          // Text-only models never receive inspection images
+          if (!chatModelSupportsVision(modelId)) {
+            return { type: 'text' as const, value: output.message };
+          }
+
+          // Local vision models rely on the hydration path for the actual image.
+          // First check storage prescence to determine output message.
+          let imageAttached = output.inspection?.imageAttached === true;
+          if (
+            providerFor(modelId) === 'local' &&
+            chatModelSupportsVision(modelId)
+          ) {
+            imageAttached = await storageObjectExists(
+              supabaseClient,
+              'images',
+              inspectionPreviewStoragePath(
+                conversation.user_id,
+                conversation.id,
+                toolCallId,
+              ),
+            );
+          }
+          const text = `${output.message}\nRendered inspection views: ${views}.\nMulti-view inspection image attached: ${imageAttached ? 'yes' : 'no'}.`;
+          return { type: 'text' as const, value: text };
+        }
       },
     },
     answer_user: chatTools.answer_user,
   };
 }
 
+function getAuxLanguageModel(providers: ChatProviders): LanguageModel | null {
+  const { activeCatalog, hasAnthropicApiKey } = getLocalChatState();
+  const auxLocal = getAuxiliaryLocalModel(activeCatalog);
+  if (auxLocal?.baseUrl) {
+    const baseUrl =
+      normalizeOpenRouterBaseUrl(auxLocal.baseUrl) ?? auxLocal.baseUrl;
+    return providers
+      .local(baseUrl, localApiKeyForModel(auxLocal))
+      .chat(auxLocal.id);
+  }
+  if (hasAnthropicApiKey) return providers.anthropic()('claude-haiku-4-5');
+  return null;
+}
+
 function chatModel(conversation: ConversationAccess, model: Model) {
   if (conversation.type === 'creative') {
     return 'anthropic/claude-sonnet-4.5';
@@ -1004,33 +1257,44 @@ export async function handleAiChatRequest(req: Request) {
     );
   }
 
-  // Pre-flight balance gate. A chat costs at least 1 billing token, so a
-  // total of 0 means we cannot let the stream start. We don't try to
+  const actualModelId = chatModel(conversation, rawBody.model);
+  if (isMissingLocalBaseUrl(actualModelId)) {
+    return jsonResponse(
+      { error: 'baseUrl is not configured for local model' },
+      503,
+    );
+  }
+
+  // Pre-flight balance gate for billable (cloud) models. Local-catalog
+  // models are billing-exempt. Cloud chat costs at least 1 billing token,
+  // so a total of 0 means we cannot let the stream start. We don't try to
   // estimate the exact cost up front — chat is variable, and the billing
   // service drains the remainder to zero if the actual usage exceeds
   // what's left (see onFinish below).
-  try {
-    const status = await billing.getStatus(user.email);
-    if (status.tokens.total <= 0) {
-      return jsonResponse(
-        {
-          error: 'insufficient_tokens',
-          code: 'insufficient_tokens',
-          tokensRequired: 1,
-          tokensAvailable: 0,
-        },
-        402,
-      );
+  if (!isActiveLocalModel(actualModelId)) {
+    try {
+      const status = await billing.getStatus(user.email);
+      if (status.tokens.total <= 0) {
+        return jsonResponse(
+          {
+            error: 'insufficient_tokens',
+            code: 'insufficient_tokens',
+            tokensRequired: 1,
+            tokensAvailable: 0,
+          },
+          402,
+        );
+      }
+    } catch (error) {
+      logError(error, {
+        functionName: 'ai-chat',
+        statusCode: error instanceof BillingClientError ? error.status : 502,
+        userId: user.id,
+        conversationId: conversation.id,
+        additionalContext: { operation: 'billing_preflight' },
+      });
+      return jsonResponse({ error: 'Billing service unavailable' }, 503);
     }
-  } catch (error) {
-    logError(error, {
-      functionName: 'ai-chat',
-      statusCode: error instanceof BillingClientError ? error.status : 502,
-      userId: user.id,
-      conversationId: conversation.id,
-      additionalContext: { operation: 'billing_preflight' },
-    });
-    return jsonResponse({ error: 'Billing service unavailable' }, 503);
   }
 
   const tools =
@@ -1038,8 +1302,8 @@ export async function handleAiChatRequest(req: Request) {
       ? creativeTools({ conversation, req, model: rawBody.model })
       : parametricTools({
           supabaseClient,
-          previewPathForToolCall: (toolCallId) =>
-            `${user.id}/${conversation.id}/inspection-preview-${toolCallId}`,
+          conversation,
+          modelId: actualModelId,
         });
 
   let branchMessages: AppUIMessage[];
@@ -1088,6 +1352,9 @@ export async function handleAiChatRequest(req: Request) {
   // the very first user turn.
   const isFirstUserTurn = branchMessages.length === 1 && leafRole === 'user';
 
+  // Resolve the actual model ID the request will run against.
+  const resolvedProvider = providerFor(actualModelId);
+
   // Rehydrate image file parts before handing them to the model. The
   // persisted `url` is a storage reference (or, for the oldest backfilled
   // rows, a dead `/public/` path), neither of which the provider can fetch —
@@ -1097,38 +1364,18 @@ export async function handleAiChatRequest(req: Request) {
   // (legacy rows that inlined base64) pass through untouched; anything we
   // can't resolve is dropped so a missing image never poisons the request
   // with an unfetchable URL.
-  const hydratedMessages = await Promise.all(
-    branchMessages.map(async (message) => ({
-      ...message,
-      parts: (
-        await Promise.all(
-          message.parts.map(async (part) => {
-            if (
-              part.type !== 'file' ||
-              typeof part.mediaType !== 'string' ||
-              !part.mediaType.startsWith('image/') ||
-              part.url.startsWith('data:')
-            ) {
-              return part;
-            }
-            const imageId = imageIdFromFilename(part.filename);
-            if (!imageId) return null;
-            const downloaded = await downloadAsBase64(
-              supabaseClient,
-              'images',
-              imageStoragePath(conversation.user_id, conversation.id, imageId),
-            );
-            if (!downloaded) return null;
-            return {
-              ...part,
-              mediaType: downloaded.mediaType,
-              url: `data:${downloaded.mediaType};base64,${downloaded.base64}`,
-            };
-          }),
-        )
-      ).filter((part): part is NonNullable<typeof part> => part != null),
-    })),
-  );
+  //
+  // For local providers on parametric conversations, strict OpenAI-compatible
+  // servers reject images inside `role: tool` messages. We instead insert a
+  // follow-up `role: user` message carrying the latest inspection render after
+  // the most recent build tool call so `convertToModelMessages` never needs to
+  // put images in tool output.
+  const hydratedMessages = await buildHydratedMessages(branchMessages, {
+    supabaseClient,
+    conversation,
+    modelId: actualModelId,
+    provider: resolvedProvider,
+  });
 
   const modelMessages = await convertToModelMessages<AppUIMessage>(
     hydratedMessages,
@@ -1161,13 +1408,6 @@ export async function handleAiChatRequest(req: Request) {
       },
     },
   );
-
-  // Resolve the actual model ID the request will run against. For
-  // `creative` conversations this is hardcoded to Sonnet regardless of
-  // what the client picked — billing has to price the model that ran,
-  // not the one the user requested.
-  const actualModelId = chatModel(conversation, rawBody.model);
-  const resolvedProvider = providerFor(actualModelId);
   const baseLogContext = {
     userId: user.id,
     conversationId: conversation.id,
@@ -1246,14 +1486,12 @@ export async function handleAiChatRequest(req: Request) {
         leafRole === 'user' &&
         stepNumber === 0
       ) {
-        // Restrict the toolset to the build tool on the first step. Models that
-        // accept a forced tool_choice get it pinned; the reasoning-tier Claude 5
-        // models (Fable/Mythos) reject forced tool use and fall back to auto,
-        // relying on the system prompt to call build_parametric_model.
-        // When pinning the tool on a thinking-enabled Anthropic model, thinking
-        // must be off for this step (Anthropic rejects forced tool use while
-        // thinking is on) — disable it here only; later steps keep the adaptive
-        // thinking configured in buildChatModel.
+        // Restrict the toolset to the build tool on the first step. Turns that
+        // accept a forced tool_choice get it pinned; turns that reject forced
+        // tool use (Claude Fable/Mythos, local models with
+        // supportsForcedToolChoice: false, or the supportsThinking fallback)
+        // fall back to auto and rely on the system prompt to call
+        // build_parametric_model.
         return {
           activeTools: ['build_parametric_model' as never],
           ...(forceBuildToolChoice
@@ -1373,14 +1611,17 @@ export async function handleAiChatRequest(req: Request) {
     execute: async ({ writer }) => {
       // Title (first user turn only) runs in parallel with the model
       // stream — fire-and-forget; the assistant doesn't wait on it.
-      if (isFirstUserTurn && env('ANTHROPIC_API_KEY')) {
-        void emitConversationTitle({
-          writer,
-          anthropic: providers.anthropic(),
-          supabaseClient,
-          conversation,
-          firstMessage: branchMessages[0],
-        });
+      if (isFirstUserTurn) {
+        const auxModel = getAuxLanguageModel(providers);
+        if (auxModel) {
+          void emitConversationTitle({
+            writer,
+            auxModel,
+            supabaseClient,
+            conversation,
+            firstMessage: branchMessages[0],
+          });
+        }
       }
 
       writer.merge(
@@ -1488,29 +1729,31 @@ export async function handleAiChatRequest(req: Request) {
               });
             }
 
-            try {
-              // Drains the user's remaining balance to zero if the
-              // request cost more than they had. The billing service
-              // accepts the partial deduction, writes an audit row as
-              // `<operation>_partial`, and the pre-flight gate above
-              // will block the next request. Not an error path —
-              // intentional terminal state. Runs after the persist above so
-              // its latency never delays the row the client is waiting on.
-              await billing.consume(user.email!, {
-                tokens: billingTokens,
-                operation:
-                  conversation.type === 'creative' ? 'chat' : 'parametric',
-                referenceId: responseMessage.id,
-              });
-            } catch (error) {
-              logError(error, {
-                functionName: 'ai-chat',
-                statusCode:
-                  error instanceof BillingClientError ? error.status : 502,
-                userId: user.id,
-                conversationId: conversation.id,
-                additionalContext: { operation: 'billing_consume' },
-              });
+            if (billingTokens > 0) {
+              try {
+                // Drains the user's remaining balance to zero if the
+                // request cost more than they had. The billing service
+                // accepts the partial deduction, writes an audit row as
+                // `<operation>_partial`, and the pre-flight gate above
+                // will block the next request. Not an error path —
+                // intentional terminal state. Runs after the persist above so
+                // its latency never delays the row the client is waiting on.
+                await billing.consume(user.email!, {
+                  tokens: billingTokens,
+                  operation:
+                    conversation.type === 'creative' ? 'chat' : 'parametric',
+                  referenceId: responseMessage.id,
+                });
+              } catch (error) {
+                logError(error, {
+                  functionName: 'ai-chat',
+                  statusCode:
+                    error instanceof BillingClientError ? error.status : 502,
+                  userId: user.id,
+                  conversationId: conversation.id,
+                  additionalContext: { operation: 'billing_consume' },
+                });
+              }
             }
 
             // Only generate suggestions once the assistant has actually
@@ -1519,28 +1762,31 @@ export async function handleAiChatRequest(req: Request) {
             // continuation `onFinish` will fire suggestions for the real
             // final state. Avoids a wasted Haiku call AND prevents
             // mid-turn placeholder pills.
-            if (!hasPendingToolCall && env('ANTHROPIC_API_KEY')) {
-              // MUST be awaited (not `void`). `createUIMessageStream`
-              // closes the SSE controller as soon as the merged stream
-              // drains — and the merged stream resolves once this
-              // `onFinish` returns. A fire-and-forget here would race
-              // the close, and the `writer.write` inside
-              // `emitConversationSuggestions` would silently no-op
-              // because `safeEnqueue` swallows enqueue errors on a
-              // closed controller (see ai/dist/index.mjs:8264). The
-              // ~200-500ms Haiku call delays the client's "streaming"
-              // → "ready" transition by the same amount, which is the
-              // tradeoff for getting pills delivered.
-              await emitConversationSuggestions({
-                writer,
-                anthropic: providers.anthropic(),
-                supabaseClient,
-                conversation,
-                branch: [
-                  ...branchMessages,
-                  { ...responseMessage, parts: finalizedParts },
-                ],
-              });
+            if (!hasPendingToolCall) {
+              const auxModel = getAuxLanguageModel(providers);
+              if (auxModel) {
+                // MUST be awaited (not `void`). `createUIMessageStream`
+                // closes the SSE controller as soon as the merged stream
+                // drains — and the merged stream resolves once this
+                // `onFinish` returns. A fire-and-forget here would race
+                // the close, and the `writer.write` inside
+                // `emitConversationSuggestions` would silently no-op
+                // because `safeEnqueue` swallows enqueue errors on a
+                // closed controller (see ai/dist/index.mjs:8264). The
+                // ~200-500ms aux-model call delays the client's "streaming"
+                // → "ready" transition by the same amount, which is the
+                // tradeoff for getting pills delivered.
+                await emitConversationSuggestions({
+                  writer,
+                  auxModel,
+                  supabaseClient,
+                  conversation,
+                  branch: [
+                    ...branchMessages,
+                    { ...responseMessage, parts: finalizedParts },
+                  ],
+                });
+              }
             }
           },
         }),
@@ -1565,19 +1811,19 @@ export async function handleAiChatRequest(req: Request) {
  */
 async function emitConversationTitle({
   writer,
-  anthropic,
+  auxModel,
   supabaseClient,
   conversation,
   firstMessage,
 }: {
   writer: UIMessageStreamWriter<AppUIMessage>;
-  anthropic: AnthropicProvider;
+  auxModel: LanguageModel;
   supabaseClient: SupabaseAnon;
   conversation: ConversationAccess;
   firstMessage: AppUIMessage;
 }) {
   try {
-    const title = await generateConversationTitle({ anthropic, firstMessage });
+    const title = await generateConversationTitle({ auxModel, firstMessage });
     await supabaseClient
       .from('conversations')
       .update({ title })
@@ -1607,20 +1853,20 @@ async function emitConversationTitle({
  */
 async function emitConversationSuggestions({
   writer,
-  anthropic,
+  auxModel,
   supabaseClient,
   conversation,
   branch,
 }: {
   writer: UIMessageStreamWriter<AppUIMessage>;
-  anthropic: AnthropicProvider;
+  auxModel: LanguageModel;
   supabaseClient: SupabaseAnon;
   conversation: ConversationAccess;
   branch: AppUIMessage[];
 }) {
   try {
     const suggestions = await generateConversationSuggestions({
-      anthropic,
+      auxModel,
       branch,
       conversationType: conversation.type,
     });
diff --git a/src/server/localChatConfig.ts b/src/server/localChatConfig.ts
new file mode 100644
index 00000000..18fb9c57
--- /dev/null
+++ b/src/server/localChatConfig.ts
@@ -0,0 +1,108 @@
+/**
+ * Server-side local-chat configuration: reads `local-models.json` from the
+ * project root. A model is active when it has a `baseUrl` set in the catalog.
+ * Pure routing logic lives in `@shared/localModels`; this module wires
+ * disk I/O for `aiChat`.
+ */
+import fs from 'node:fs';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+import {
+  isMissingLocalBaseUrl as isMissingLocalBaseUrlCore,
+  isLocalModelId,
+  parseLocalModelsJson,
+  providerFor as providerForCore,
+  type ChatProvider,
+  type LocalModelConfig,
+} from '@shared/localModels';
+import { env } from './env';
+
+export type { ChatProvider };
+
+export type LocalChatState = {
+  enabled: boolean;
+  /** On-disk catalog, even when local chat is disabled. */
+  catalog: LocalModelConfig[];
+  /** Models with a configured baseUrl. */
+  activeCatalog: LocalModelConfig[];
+  hasAnthropicApiKey: boolean;
+};
+
+const projectRoot = path.resolve(
+  path.dirname(fileURLToPath(import.meta.url)),
+  '../..',
+);
+
+// Cached after first read; restart the dev server after editing `local-models.json`.
+let cachedCatalog: LocalModelConfig[] | undefined;
+
+function loadCatalogFromDisk(): LocalModelConfig[] {
+  if (cachedCatalog) return cachedCatalog;
+
+  const userPath = path.join(projectRoot, 'local-models.json');
+  if (!fs.existsSync(userPath)) {
+    cachedCatalog = [];
+    return cachedCatalog;
+  }
+
+  try {
+    const raw = JSON.parse(fs.readFileSync(userPath, 'utf8')) as unknown;
+    cachedCatalog = parseLocalModelsJson(raw);
+  } catch (error) {
+    console.error(
+      `Failed to load local-models.json from ${userPath}; using empty catalog.`,
+      error,
+    );
+    cachedCatalog = [];
+  }
+
+  return cachedCatalog;
+}
+
+export function getLocalChatState(): LocalChatState {
+  const catalog = loadCatalogFromDisk();
+  const activeCatalog = catalog.filter((m) => m.baseUrl?.trim());
+
+  return {
+    enabled: activeCatalog.length > 0,
+    catalog,
+    activeCatalog,
+    hasAnthropicApiKey: env('ANTHROPIC_API_KEY').trim().length > 0,
+  };
+}
+
+export function providerFor(modelId: string): ChatProvider {
+  const { activeCatalog } = getLocalChatState();
+  return providerForCore(modelId, activeCatalog);
+}
+
+export function isActiveLocalModel(modelId: string): boolean {
+  return isLocalModelId(modelId, getLocalChatState().activeCatalog);
+}
+
+export function getActiveLocalModel(
+  modelId: string,
+): LocalModelConfig | undefined {
+  return getLocalChatState().activeCatalog.find(
+    (model) => model.id === modelId,
+  );
+}
+
+// Resolve local API key from env without exposing secrets to the client bundle.
+// If `apiKey` is set in `local-models.json`, it is interpreted as the env var
+// name (for example `DEEPSEEK_API_KEY`), not a raw secret.
+export function localApiKeyForModel(
+  model: LocalModelConfig,
+): string | undefined {
+  const configuredName = model.apiKey?.trim();
+  if (!configuredName) return undefined;
+  const configured = env(configuredName).trim();
+  return configured || undefined;
+}
+
+// Client may show catalog ids whose baseUrl is missing on the server — reject
+// those requests with a clear error.
+export function isMissingLocalBaseUrl(modelId: string): boolean {
+  const { catalog } = getLocalChatState();
+  return isMissingLocalBaseUrlCore(modelId, catalog);
+}
diff --git a/src/server/messageUtils.test.ts b/src/server/messageUtils.test.ts
new file mode 100644
index 00000000..e6da449a
--- /dev/null
+++ b/src/server/messageUtils.test.ts
@@ -0,0 +1,90 @@
+import assert from 'node:assert/strict';
+import { describe, it } from 'node:test';
+import type { AppUIMessage } from '@shared/chatAi';
+import { findLatestBuildToolCallId } from './messageUtils.ts';
+
+function buildToolPart(
+  toolCallId: string,
+  state:
+    | 'output-available'
+    | 'input-available'
+    | 'output-error' = 'output-available',
+): AppUIMessage['parts'][number] {
+  return {
+    type: 'tool-build_parametric_model',
+    toolCallId,
+    state,
+    input: { code: 'cube(1);' },
+    ...(state === 'output-available'
+      ? {
+          output: {
+            status: 'success' as const,
+            message: 'ok',
+            inspection: { views: ['ISO'] as const, imageAttached: true },
+          },
+        }
+      : state === 'output-error'
+        ? { errorText: 'failed' }
+        : {}),
+  } as AppUIMessage['parts'][number];
+}
+
+describe('findLatestBuildToolCallId', () => {
+  it('returns null for an empty branch', () => {
+    assert.equal(findLatestBuildToolCallId([]), null);
+  });
+
+  it('returns the only completed build tool call', () => {
+    const messages: AppUIMessage[] = [
+      {
+        id: 'a1',
+        role: 'assistant',
+        parts: [buildToolPart('call-1')],
+      },
+    ];
+    assert.equal(findLatestBuildToolCallId(messages), 'call-1');
+  });
+
+  it('returns the latest build across assistant messages', () => {
+    const messages: AppUIMessage[] = [
+      {
+        id: 'a1',
+        role: 'assistant',
+        parts: [buildToolPart('call-1')],
+      },
+      { id: 'u1', role: 'user', parts: [{ type: 'text', text: 'taller' }] },
+      {
+        id: 'a2',
+        role: 'assistant',
+        parts: [buildToolPart('call-2')],
+      },
+    ];
+    assert.equal(findLatestBuildToolCallId(messages), 'call-2');
+  });
+
+  it('returns the last completed build part within one assistant message', () => {
+    const messages: AppUIMessage[] = [
+      {
+        id: 'a1',
+        role: 'assistant',
+        parts: [buildToolPart('call-1'), buildToolPart('call-2')],
+      },
+    ];
+    assert.equal(findLatestBuildToolCallId(messages), 'call-2');
+  });
+
+  it('ignores non-output build tool parts', () => {
+    const messages: AppUIMessage[] = [
+      {
+        id: 'a1',
+        role: 'assistant',
+        parts: [
+          buildToolPart('call-error', 'output-error'),
+          buildToolPart('call-pending', 'input-available'),
+          buildToolPart('call-done'),
+        ],
+      },
+    ];
+    assert.equal(findLatestBuildToolCallId(messages), 'call-done');
+  });
+});
diff --git a/src/server/messageUtils.ts b/src/server/messageUtils.ts
index adda5a40..0df23eed 100644
--- a/src/server/messageUtils.ts
+++ b/src/server/messageUtils.ts
@@ -1,4 +1,5 @@
-import { env } from './env';
+import { env } from './env.ts';
+import type { AppUIMessage } from '@shared/chatAi';
 
 export function reformatSignedUrl(signedUrl: string): string {
   const supabaseHost = (
@@ -8,3 +9,33 @@ export function reformatSignedUrl(signedUrl: string): string {
   const url = new URL(signedUrl);
   return `${supabaseHost}${url.pathname}${url.search}`;
 }
+
+/**
+ * Latest `build_parametric_model` with `output-available` on the branch.
+ * `buildHydratedMessages` uses this so local parametric only re-downloads
+ * one inspection preview per request (newest-to-oldest walk).
+ */
+export function findLatestBuildToolCallId(
+  messages: AppUIMessage[],
+): string | null {
+  for (let index = messages.length - 1; index >= 0; index--) {
+    const message = messages[index];
+    if (message.role !== 'assistant') continue;
+    for (
+      let partIndex = message.parts.length - 1;
+      partIndex >= 0;
+      partIndex--
+    ) {
+      const part = message.parts[partIndex];
+      if (
+        part.type === 'tool-build_parametric_model' &&
+        part.state === 'output-available' &&
+        'toolCallId' in part &&
+        typeof part.toolCallId === 'string'
+      ) {
+        return part.toolCallId;
+      }
+    }
+  }
+  return null;
+}
diff --git a/src/views/EditorView.tsx b/src/views/EditorView.tsx
index bed7cd7d..d6c4f4e6 100644
--- a/src/views/EditorView.tsx
+++ b/src/views/EditorView.tsx
@@ -14,6 +14,7 @@ import { OpenSCADPreview } from '@/components/viewer/OpenSCADViewer';
 import { MeshPreview } from '@/components/viewer/MeshPreview';
 import Loader from '@/components/viewer/Loader';
 import { useAuth } from '@/contexts/AuthContext';
+import { useIsLocalModel } from '@/hooks/useLocalModels.ts';
 import { ConversationContext } from '@/contexts/ConversationContext';
 import { SelectedItemsContext } from '@/contexts/SelectedItemsContext';
 import { useConversation } from '@/contexts/ConversationContext';
@@ -194,6 +195,7 @@ function ConversationEditor() {
         ? 'quality'
         : 'google/gemini-3.1-pro-preview'),
   );
+  const isLocalModel = useIsLocalModel(model);
   const [activePreview, setActivePreview] = useState<ActivePreview>(null);
   const [parameters, setParameters] = useState<Parameter[]>([]);
   const [currentOutput, setCurrentOutput] = useState<Blob | undefined>();
@@ -705,7 +707,7 @@ function ConversationEditor() {
             initialBranch={initialBranch}
             model={model}
             setModel={updateSelectedModel}
-            isDisabled={totalTokens <= 0}
+            isDisabled={totalTokens <= 0 && !isLocalModel}
             onSendParts={handleSendParts}
             onRetry={handleRetry}
             onEdit={handleEdit}
diff --git a/src/views/PromptView.tsx b/src/views/PromptView.tsx
index 1d5d7f95..9a54ce9f 100644
--- a/src/views/PromptView.tsx
+++ b/src/views/PromptView.tsx
@@ -14,6 +14,7 @@ import { LowPromptsWarningMessage } from '@/components/LowPromptsWarningMessage'
 import { NewProductBanner } from '@/components/NewProductBanner';
 import { FreePlanTrialPill } from '@/components/FreePlanTrialPill';
 import { useIsMobile } from '@/hooks/useIsMobile';
+import { useIsLocalModel } from '@/hooks/useLocalModels.ts';
 import { cn } from '@/lib/utils';
 import { SelectedItemsContext } from '@/contexts/SelectedItemsContext';
 import posthog from 'posthog-js';
@@ -65,6 +66,8 @@ export function PromptView() {
 
   const [model, setModel] = useState<Model>('google/gemini-3.1-pro-preview');
 
+  const isLocalModel = useIsLocalModel(model);
+
   const handleTypeChange = (newType: 'parametric' | 'creative') => {
     setType(newType);
     // Reset model to the default for the new type
@@ -91,8 +94,9 @@ export function PromptView() {
 
   const limitReached = useMemo(() => {
     if (isLoading) return false;
+    if (type === 'parametric' && isLocalModel) return false;
     return totalTokens <= 0;
-  }, [totalTokens, isLoading]);
+  }, [totalTokens, isLoading, type, isLocalModel]);
 
   // Trigger fade in on mount
   useEffect(() => {