feat: add scoped MCP destinations and composable CLI help

.claude-plugin/marketplace.json

@@ -7,6 +7,11 @@
     "url": "https://allagents.dev"
   },
   "plugins": [
+    {
+      "name": "allagents",
+      "description": "End-user guidance for managing AllAgents through its CLI",
+      "source": "./plugins/allagents"
+    },
     {
       "name": "deepwiki",
       "description": "AI-generated documentation for GitHub repositories",
@@ -16,6 +21,11 @@
       "name": "engineering",
       "description": "Reusable engineering workflow skills",
       "source": "./plugins/engineering"
+    },
+    {
+      "name": "tradingview",
+      "description": "Official TradingView market data and analytics",
+      "source": "./plugins/tradingview"
     }
   ]
 }

.github/plugin/marketplace.json

@@ -7,10 +7,20 @@
     "url": "https://allagents.dev"
   },
   "plugins": [
+    {
+      "name": "allagents",
+      "description": "End-user guidance for managing AllAgents through its CLI",
+      "source": "./plugins/allagents"
+    },
     {
       "name": "deepwiki",
       "description": "AI-generated documentation for GitHub repositories",
       "source": "./plugins/deepwiki"
+    },
+    {
+      "name": "tradingview",
+      "description": "Official TradingView market data and analytics",
+      "source": "./plugins/tradingview"
     }
   ]
 }

AGENTS.md

@@ -184,7 +184,4 @@ bun run build
 
 ## Interactive Testing
 
-### agent-tui
-- Use `agent-tui` when you need to exercise interactive terminal behavior manually.
-- Prefer testing the built CLI inside a temporary workspace so the interaction matches real user conditions.
-- When documenting manual verification, record the exact command, the temp workspace setup, and what terminal behavior you confirmed.
+- TUI navigation, prompt, copy, status, or recovery changes: use the installed `agent-tui` skill for terminal automation, then apply the AllAgents-specific acceptance criteria in [`docs/agent-guides/tui-dogfooding.md`](docs/agent-guides/tui-dogfooding.md) before implementation and final verification.

CHANGELOG.md

@@ -4,9 +4,13 @@
 
 ### Breaking Changes
 
-- **MCP proxy command**: Removed the temporary `allagents mcp proxy-stdio` alias. Use `allagents mcp proxy <serverUrl>` instead.
+- **MCP setup commands**: HTTP servers added with `allagents mcp add` now
+  authenticate and route through AllAgents automatically. The public
+  `--proxy` option and `mcp auth <serverUrl>` command were replaced by
+  `mcp reauth <name>`. The generated `mcp proxy` helper remains internal.
 
-  **Migration**: Re-run `allagents mcp update` or `allagents update` after upgrading so synced client configs are regenerated with `mcp proxy`.
+  **Migration**: Remove `--proxy` from `mcp add` calls. Replace
+  `allagents mcp auth <url>` with `allagents mcp reauth <configured-name>`.
 
 - **Plugin Git ref terminology**: Renamed workspace plugin `pin` to `ref`, CLI
   `--pin` to `--ref`, and sync-state `pinnedRef` to `requestedRef`. Inline
@@ -15,8 +19,38 @@
   **Migration**: Replace `pin:` with `ref:` in plugin objects and `--pin` with
   `--ref` in scripts. This is a clean cutover; the old names are not accepted.
 
+### Fixed
+
+- Project-scoped Copilot MCP servers are now written to `.github/mcp.json`,
+  which Copilot CLI discovers, instead of the unsupported
+  `.copilot/mcp-config.json` project path.
+- Interactive OAuth guidance for `mcp add` and `mcp reauth` now uses normal
+  terminal output instead of the error channel. Callback URLs are entered
+  through an abortable masked prompt, and failed reauthentication restores the
+  previous working credentials.
+
 ### Added
 
+- Added a thin first-party AllAgents skill that treats the installed CLI as
+  authoritative and follows its progressive `--help --json` indexes and leaf
+  contracts instead of relying on memorized commands.
+- Added progressive machine-readable CLI help: concise root and group indexes
+  lead to leaf contracts with usage guidance, interaction requirements,
+  expected output, options, examples, and output schemas.
+
+- Added the official TradingView MCP plugin with OAuth-backed access to market
+  data, analytics, watchlists, alerts, news, and screeners.
+- Added automatic OAuth login to `allagents mcp add` and named credential
+  renewal with `allagents mcp reauth`, including local loopback completion and
+  remote callback URL paste with strict redirect and state validation.
+- Generated HTTP MCP bridges now invoke the current pinned AllAgents version
+  through cached `npx`, so managed MCP connections do not require a global
+  AllAgents installation.
+- Full MCP server management in the interactive TUI, including destination
+  selection, listing, inspection, add, reauthenticate, and remove flows for
+  project, user, and named-profile declarations. Client configuration updates
+  automatically after mutations, with a contextual retry when an update fails.
+
 - Pi and OMP as file-sync clients at project and user scope, including native
   runtime skill paths and agent instructions.
 - Native Pi package and OMP marketplace-plugin lifecycle support for install,

README.md

@@ -24,6 +24,13 @@ npx allagents update
 
 No cloning required — AllAgents fetches the `workspace.yaml` directly from GitHub and sets up everything.
 
+Install the end-user AllAgents skill when you want your coding agent to operate
+the CLI using its current machine-readable command contracts:
+
+```bash
+npx allagents plugin install allagents@allagentsdev/allagents --scope user
+```
+
 ## How It Works
 
 1. **Configure** your workspace with repos, plugins, and target clients in `workspace.yaml`
@@ -101,9 +108,9 @@ clients:
 | `allagents profile remove <name> --yes` | Remove unchanged managed profile resources |
 | `allagents skill add <name> [--from <source>] [--scope project\|user] [--client <clients>] [--yes]` | Add source-backed skills for selected scope and clients (plural `skills` alias supported) |
 | `allagents skill list` | List skills and status |
-| `allagents mcp add <name> <commandOrUrl>` | Add an MCP server and sync to clients |
-| `allagents mcp proxy <serverUrl>` | Bridge a remote HTTP MCP server to local stdio |
-| `allagents mcp list` | List workspace MCP servers |
+| `allagents mcp add <name> <commandOrUrl> [--scope user \| --profile <name>]` | Add, authenticate, and sync an MCP server |
+| `allagents mcp reauth <name> [--scope user \| --profile <name>]` | Reauthenticate an HTTP MCP server in one destination |
+| `allagents mcp list [--scope user \| --profile <name>]` | List MCP declarations in one destination |
 | `allagents workspace status` | Show workspace state |
 | `allagents self update` | Update AllAgents CLI |
 

docs/agent-guides/tui-dogfooding.md

@@ -0,0 +1,78 @@
+# TUI Dogfooding
+
+This guide defines AllAgents-specific acceptance criteria for changes to interactive CLI navigation, prompts, labels, status messages, or failure recovery.
+
+## Tooling boundary
+
+Use the installed `agent-tui` skill for terminal automation. That skill owns installation checks, command selection, session lifecycle, snapshots, actions, waits, assertions, and cleanup. Follow its current CLI workflow rather than reproducing command recipes here.
+
+This guide owns the product-specific UX questions, interaction coverage, and evidence required for AllAgents.
+
+## Goal
+
+Prove that the built TUI is understandable to a first-time user, not only that its underlying mutation succeeds.
+
+## Setup
+
+1. Build the CLI.
+2. Create isolated temporary project and HOME directories.
+3. Seed the smallest realistic configuration that exposes every changed state.
+4. Use the `agent-tui` skill to drive the built CLI and capture each changed decision screen before interacting with it.
+
+Never dogfood against a real user workspace when an isolated fixture can exercise the behavior.
+
+## Confusion pass
+
+For every changed screen, state:
+
+- the object the user is currently managing;
+- the single decision the screen asks them to make;
+- where Back and Ctrl+C will land;
+- whether the selected scope or destination remains visible and intact.
+
+Apply one screen, one decision:
+
+- A resource list contains resources plus Add and Back.
+- A resource detail contains actions for that resource plus Back.
+- Scope or destination changes happen by returning to the chooser.
+- Maintenance mechanics stay automatic. Show a retry only when automatic recovery fails.
+
+If a menu mixes resource selection, navigation, and maintenance operations, simplify it before continuing.
+
+## Language pass
+
+Use the established public term for each operation in labels, progress messages, results, errors, and documentation. Internal implementation terms do not belong in user-facing copy. For example, use **Update** rather than sync or reconcile.
+
+Read the complete screen, not only the changed label. Adjacent hints, summaries, and success or error messages must use the same vocabulary.
+
+## Interaction pass
+
+Exercise every changed path that applies:
+
+1. Enter the flow from the main menu.
+2. Move forward through each chooser and detail screen.
+3. Use Back from every changed level.
+4. Use Ctrl+C from every changed level.
+5. Complete a successful mutation and verify both the next screen and filesystem result.
+6. Trigger a realistic failure and verify the error leaves a clear recovery path.
+7. Exercise retry, repeated retry failure, explicit cancellation, and eventual success when retry behavior changed.
+
+A transition passes only when it lands on the screen a user would predict without losing or silently changing scope.
+
+## Durable coverage
+
+Keep regression tests for navigation state, scope preservation, cancellation, mutation boundaries, and failure recovery. Test exact copy only when the wording is a deliberate product contract; use the manual confusion pass for general prose quality.
+
+## Completion evidence
+
+Record in the PR description:
+
+- the exact built command;
+- temporary workspace and HOME setup;
+- selections and transitions exercised;
+- screenshots or the text of each observed decision screen;
+- resulting configuration or filesystem state;
+- failure and retry behavior checked;
+- cleanup performed.
+
+Dogfooding is complete only when the full changed journey passes the confusion, language, interaction, and filesystem checks.

docs/plans/2026-02-03-workspace-sync-both-scopes.md

@@ -446,7 +446,7 @@ In `src/cli/metadata/workspace.ts`, remove the `--scope` example and option:
 
 **Step 2: Run help text tests**
 
-Run: `bun test tests/e2e/cli-help.test.ts tests/e2e/cli-enriched-help.test.ts tests/e2e/cli-agent-help.test.ts`
+Run: `bun test tests/e2e/cli-help.test.ts tests/e2e/cli-enriched-help.test.ts tests/unit/cli/structured-help.test.ts`
 Expected: PASS (help tests should not hardcode --scope for sync)
 
 **Step 3: Commit**

docs/plans/2026-08-10-001-feat-skill-update-deletion-handling-plan.md

@@ -222,7 +222,7 @@ flowchart TB
 - **Goal:** Expose the orchestration as a predictable interactive and scriptable command.
 - **Requirements:** R1-R3, R6, R8-R12; F1-F3.
 - **Dependencies:** U1, U2.
-- **Files:** `src/cli/commands/plugin-skills.ts`, `src/cli/metadata/plugin-skills.ts`, `src/cli/skill-arg-normalizer.ts`, `tests/unit/cli/skill-update.test.ts`, `tests/unit/cli/agent-help.test.ts`.
+- **Files:** `src/cli/commands/plugin-skills.ts`, `src/cli/metadata/plugin-skills.ts`, `src/cli/skill-arg-normalizer.ts`, `tests/unit/cli/skill-update.test.ts`, `tests/unit/cli/structured-help.test.ts`.
 - **Approach:** Register `update`; validate scope/filter values; use Clack for scope and confirmation prompts; group warnings once per physical refresh unit with every impacted scope/plugin listed; collect all decisions before execution; interpret No as retain/skip and cancel as pre-mutation abort; render one concise summary; produce the same result model and exit contract through JSON without UI noise.
 - **Patterns to follow:** Existing skill search scope picker, global JSON envelope helpers, and enriched command metadata.
 - **Test scenarios:**

docs/plans/2026-09-17-1211-perf-update-no-op-paths-plan.md

@@ -188,7 +188,7 @@ flowchart LR
 
 ### System-Wide Impact
 
-- **CLI and agents:** Human output, `--json`, exit codes, and agent-help schemas stay stable. Automation still sees successful no-op checks as existing update outcomes.
+- **CLI and agents:** Human output, `--json`, exit codes, and structured-help schemas stay stable. Automation still sees successful no-op checks as existing update outcomes.
 - **Filesystem state:** Direct plugin and skill no-ops avoid persistent checkout work while retaining existing scope synchronization. Marketplace no-ops retain the registry timestamp write required by compatibility.
 - **Dependency direction:** CLI/TUI callers create a neutral context and pass it into domain updaters; domain updaters consume context/Git/identity helpers. The leaf identity helper and Git module never import plugin, marketplace, skill, or CLI/TUI modules.
 - **Scope ownership:** Remote and physical checkout facts can be shared, but each consumer independently derives public output and writes only its owning registry. External-plugin `changed` is the OR of successful marketplace and external-checkout physical changes without changing current public precedence or sync eligibility.

docs/plans/2026-09-19-1012-feat-interactive-install-scope-clients-plan.md

@@ -80,7 +80,7 @@ The feature is primarily UX. Existing marketplace resolution, declaration instal
 - A shared summary-data model rendered by CLI and TUI.
 - Localized project/user declaration persistence for client overrides while preserving object fields.
 - Crash-safe publication of the targeted project/user config via the repository’s established same-directory temporary-write-and-rename pattern.
-- Existing command metadata, agent help, user docs, changelog, focused regression tests, and built-CLI dogfood.
+- Existing command metadata, structured help, user docs, changelog, focused regression tests, and built-CLI dogfood.
 
 ### Out of Scope
 
@@ -231,7 +231,7 @@ Only these areas require code-review attention beyond dogfooding:
 - `src/cli/tui/actions/plugins.ts`
 - `src/cli/metadata/plugin.ts`
 - `tests/unit/cli/tui-plugin-install.test.ts` (new)
-- `tests/unit/cli/agent-help.test.ts`
+- `tests/unit/cli/structured-help.test.ts`
 - `tests/e2e/plugin-install-options.test.ts` (new)
 
 **Changes:**
@@ -281,7 +281,7 @@ Only these areas require code-review attention beyond dogfooding:
 
 - Document chooser order, first-config default consequence, per-plugin overrides, flags, and non-interactive behavior.
 - Add a valid `plugins[].clients` example using `source`.
-- Update command metadata and generated agent help.
+- Update command metadata and structured help.
 - Remove temporary dogfood workspaces and any obsolete plan artifacts after implementation.
 
 ## Verification Contract

docs/public/schemas/v1/project-workspace.schema.json

@@ -315,7 +315,8 @@
               "type": "array",
               "items": {
                 "type": "string"
-              }
+              },
+              "default": []
             },
             "servers": {
               "type": "object",
@@ -336,9 +337,6 @@
               }
             }
           },
-          "required": [
-            "clients"
-          ],
           "additionalProperties": true
         },
         "mcpServers": {

docs/public/schemas/v1/user-workspace.schema.json

@@ -318,7 +318,8 @@
               "type": "array",
               "items": {
                 "type": "string"
-              }
+              },
+              "default": []
             },
             "servers": {
               "type": "object",
@@ -339,9 +340,6 @@
               }
             }
           },
-          "required": [
-            "clients"
-          ],
           "additionalProperties": true
         },
         "mcpServers": {
@@ -1038,7 +1036,13 @@
                       "additionalProperties": false
                     }
                   ]
+                },
+                "propertyNames": {
+                  "pattern": "^[A-Za-z0-9_.-]{1,100}$"
                 }
+              },
+              "mcpProxy": {
+                "$ref": "#/definitions/AllAgentsUserWorkspace/properties/mcpProxy"
               }
             },
             "required": [

docs/src/content/docs/docs/getting-started/installation.mdx

@@ -22,7 +22,9 @@ npx allagents
 ```
 
 :::note
-If you plan to use the [MCP Proxy](/docs/guides/mcp-proxy/) feature, install AllAgents globally (npm or bun) instead. The proxy command is invoked directly by your MCP clients later, not by you through npx, so `allagents` needs to already be resolvable on `PATH` at that point.
+HTTP MCP servers also work when setup is run through `npx`. Generated client
+configs invoke a pinned AllAgents version through `npx`, which reuses npm's
+package cache after the first launch.
 :::
 
 ## Using bun

docs/src/content/docs/docs/guides/agent-portability.mdx

@@ -124,7 +124,7 @@ Running `allagents update` syncs agents, skills, hooks, and MCP servers to both
 | Hooks | `.claude/hooks/` | `.github/hooks/` |
 | Commands / Prompts | `.claude/commands/` | `.github/prompts/` |
 | Agent file | `CLAUDE.md` | `AGENTS.md` |
-| MCP servers | `.claude/mcp.json` | `.copilot/mcp-config.json` |
+| MCP servers | `.claude/mcp.json` | `.github/mcp.json` |
 | WORKSPACE-RULES | Injected in `CLAUDE.md` | Injected in `AGENTS.md` |
 
 ### Skills Are the Strongest Interop Point