Mission#

Wire the local Ollama instance into OpenClaw as an OpenAI-compatible provider so any agent can route specific tasks/cron jobs to a local model with --model. Verify end-to-end: openclaw agent --agent main --model "ollama/qwen2.5:7b-instruct-q4_K_M" --message "reply: ok" returns ok and the call hit the local server (not Anthropic).

Steps#

1. Read the current ~/.openclaw/openclaw.json, find agents.defaults.models.providers (or equivalent — verify schema with openclaw config schema lookup).
2. Add an ollama provider entry:
   • Type: OpenAI-compatible
   • Base URL: http://127.0.0.1:11434/v1
   • Auth: none (local)
   • Models: at least qwen2.5:7b-instruct-q4_K_M, qwen2.5-coder:7b-q4_K_M
   • Useful aliases: qwen → ollama/qwen2.5:7b-instruct-q4_K_M, qwen-coder → ollama/qwen2.5-coder:7b-q4_K_M
3. openclaw config validate — must pass
4. openclaw gateway restart (or hot-reload if config supports it; if not, the restart will kill the active session — Maxime knows this trade-off)
5. Smoke test (from a fresh shell, not a session running through the gateway):
   • openclaw agent --agent main --model qwen --message "reply: ok" → returns ok
   • During the call, watch journalctl --user -u ollama.service -f to confirm the request hit Ollama
6. Document the provider config snippet in MEMORY.md so I can re-add it if it ever gets stripped

Acceptance criteria#

• Provider ollama registered and visible in openclaw config get agents.defaults.models
• Aliases qwen and qwen-coder resolve to the right provider/model ids
• CLI smoke test succeeds and we can see the request hit Ollama logs
• No regression on existing Anthropic / OpenAI providers (run a quick Robbie turn to confirm)

Files / locations#

• ~/.openclaw/openclaw.json (live config)
• MEMORY.md update with the snippet