Code Ollama Adapter

Local Ollama Cloud proxy for Claude Code and Codex. It adapts GLM-5.2 for max reasoning, 1M-class context aliases, Anthropic-compatible requests, and OpenAI Responses API traffic.

The adapter is intentionally small: it forwards OpenAI-compatible and Anthropic-compatible HTTP traffic to a local Ollama server, rewriting only the metadata that these clients cannot express correctly.

What It Adapts

• Claude Code can select glm-5.2:cloud[1m] so its local context budget is 1M-class. The adapter forwards the valid Ollama model name glm-5.2:cloud.
• Claude Code can use a cheaper Haiku-class model such as glm-4.7:cloud for lightweight tasks while keeping Sonnet/Opus on GLM-5.2.
• Codex can expose xhigh as the user-selectable top reasoning level. The adapter forwards Ollama GLM-compatible max.
• Normal high, medium, low, and none reasoning levels are left alone.

Runtime Shape

Claude Code ─┐
             ├─ http://127.0.0.1:11435 ── http://127.0.0.1:11434 ── Ollama Cloud
Codex ───────┘

Default rewrites:

• model: "glm-5.2" -> model: "glm-5.2:cloud"
• model: "glm-5.2:cloud[1m]" -> model: "glm-5.2:cloud"
• reasoning.effort: "xhigh" -> reasoning.effort: "max" on Responses API
• reasoning_effort: "xhigh" -> reasoning_effort: "max" on Chat Completions

The service logs method, path, status, and mutation names only. It does not log prompts or response bodies.

Privacy And Security

• The default service binds to 127.0.0.1; do not expose it publicly without a separate authentication and network security plan.
• Do not commit API keys, provider tokens, .env files, local Claude/Codex credential files, prompt logs, response logs, or session transcripts.
• .gitignore excludes common local secrets, logs, build output, and temporary files.
• The optional pre-commit setup includes staged gitleaks scanning.

Layout

• cmd/code-ollama-adapter: Go binary entrypoint.
• internal/proxy: request forwarding and payload rewrite logic.
• systemd/code-ollama-adapter.service: Linux systemd unit template.
• launchd/ai.openclaw.code-ollama-adapter.plist: macOS launchd template.
• scripts/install.sh: build and install for Linux/macOS.
• scripts/uninstall.sh: stop and remove service metadata.
• config/codex-ollama-cloud.config.toml: Codex profile example.
• .agents/skills/configure-code-ollama-adapter: Agent workflow for safely configuring and verifying this local setup.
• .agents/skills/setup-codex-claude-ollama: Agent workflow for helping users configure Codex and Claude Code from scratch.

Agent Skill

This repo includes a small Codex skill for agent-assisted setup: setup-codex-claude-ollama.

Ask an agent to use $setup-codex-claude-ollama when you want it to configure Codex and Claude Code on a machine, preserve existing settings, back up config files, set Codex xhigh, set Claude's 1M-class context alias, start the adapter, and verify both clients.

Use $configure-code-ollama-adapter for narrower project maintenance and repair after the adapter is already installed. Use the README for human maintenance.

Build And Test

git clone git@github.com:hxy91819/code-ollama-adapter.git
cd code-ollama-adapter
go test ./...
go build -o bin/code-ollama-adapter ./cmd/code-ollama-adapter

CI runs on Ubuntu and macOS and checks:

• gofmt
• go test ./...
• binary build
• install/uninstall script help output

Optional local hygiene:

pre-commit install
pre-commit run --all-files

The pre-commit setup expects gitleaks and golangci-lint to be installed. It runs staged secret scanning, basic file hygiene checks, gofmt, go vet, go test, and a small Go lint set.

Run Manually

~/code-ollama-adapter/bin/code-ollama-adapter \
  --host 127.0.0.1 \
  --port 11435 \
  --upstream http://127.0.0.1:11434 \
  --model-alias glm-5.2 \
  --model-alias 'glm-5.2:cloud[1m]' \
  --model-target glm-5.2:cloud \
  --reasoning-map xhigh=max \
  --default-reasoning-effort max

Health check:

curl http://127.0.0.1:11435/health

Install As A Service

Linux systemd:

cd ~/code-ollama-adapter
scripts/install.sh
systemctl status code-ollama-adapter.service

The installer builds bin/code-ollama-adapter, installs a root-owned executable under /usr/local/lib/code-ollama-adapter/<service-name>/, and writes a systemd unit that executes that installed binary. The service does not execute a mutable checkout binary, and distinct service names get distinct binary paths.

macOS launchd:

cd ~/code-ollama-adapter
scripts/install.sh
launchctl print "gui/$(id -u)/ai.openclaw.code-ollama-adapter"

When run as root on macOS, the installer writes to /Library/LaunchDaemons; otherwise it writes to ~/Library/LaunchAgents. The launchd plist is generated with the installed binary path.

Uninstall

Linux:

cd ~/code-ollama-adapter
scripts/uninstall.sh

macOS:

cd ~/code-ollama-adapter
scripts/uninstall.sh

Uninstall removes service metadata only. It does not delete the project tree or the installed binary.

Codex Profile

codexo uses:

codex --profile ollama-cloud --dangerously-bypass-approvals-and-sandbox

The profile should point the Ollama Cloud provider at the adapter:

model = "glm-5.2"
model_provider = "ollama_cloud"
model_reasoning_effort = "xhigh"
model_catalog_json = "/absolute/path/to/ollama-cloud-glm-5.2.json"

[model_providers.ollama_cloud]
name = "Ollama Cloud via Code Ollama Adapter"
base_url = "http://127.0.0.1:11435/v1"
wire_api = "responses"

The model catalog should include xhigh in supported_reasoning_levels. This setup defaults Codex to xhigh, and the service also injects max when Codex omits reasoning from the wire request. Use high explicitly when you want normal high reasoning, and xhigh for Ollama GLM max:

codex exec --profile ollama-cloud \
  -c model_reasoning_effort='"xhigh"' \
  --dangerously-bypass-approvals-and-sandbox \
  'Reply exactly OK.'

Claude Code Setup

Claude Code should use the adapter as its Anthropic-compatible base URL and a display model with the [1m] suffix for Sonnet/Opus:

ANTHROPIC_BASE_URL=http://127.0.0.1:11435
ANTHROPIC_DEFAULT_SONNET_MODEL='glm-5.2:cloud[1m]'
ANTHROPIC_DEFAULT_HAIKU_MODEL='glm-4.7:cloud'
ANTHROPIC_DEFAULT_OPUS_MODEL='glm-5.2:cloud[1m]'

The adapter strips [1m] before forwarding to Ollama. Claude Code still budgets Sonnet/Opus sessions as 1M-class models. glm-4.7:cloud is intentionally left without [1m]; it is a cheaper Haiku-class default with a smaller context window.

Verification

Request-level smoke:

curl -sS http://127.0.0.1:11435/v1/responses \
  -H 'content-type: application/json' \
  -d '{"model":"glm-5.2","input":"Reply exactly OK.","max_output_tokens":16,"reasoning":{"effort":"xhigh"}}' \
  | jq '{model, reasoning}'

Expected result includes:

{"reasoning":{"effort":"max"}}

Client smoke:

claude -p /context --model 'glm-5.2:cloud[1m]'
claude -p 'Reply exactly OK.' --model 'glm-4.7:cloud'

codex exec --profile ollama-cloud \
  -c model_reasoning_effort='"xhigh"' \
  --dangerously-bypass-approvals-and-sandbox \
  'Reply exactly OK.'

Expected log markers:

• rewrite_model: a client alias was mapped to glm-5.2:cloud.
• rewrite_reasoning: xhigh was mapped to max.

Maintenance

• Keep rewrite rules explicit and narrow.
• Add a test in internal/proxy/transform_test.go for every new rewrite rule.
• Keep the service listening on 127.0.0.1 unless there is a concrete need to expose it.
• Do not log request or response bodies.
• After editing behavior, run go test ./..., reinstall/restart the service, and run at least one live smoke through Codex or Claude Code.

Migration Notes

Earlier prototypes used these names:

• project directory: ollama-cloud-code-proxy
• service: ollama-cloud-code-proxy.service
• binary: ollama-cloud-code-proxy
• older service: claude-ollama-proxy.service

The current canonical name is code-ollama-adapter. Keep only one service active on port 11435.