Ollama (Provider)

How it works

The /add-ollama-provider skill routes a NanoClaw agent group to a local Ollama instance instead of the Anthropic API. Because Ollama already speaks the Anthropic API natively (v1/messages), there’s no provider code to change — just an env override telling the Anthropic SDK where to send its requests.

This is a per-group switch. Other groups in the same NanoClaw install can keep talking to Claude.

Setup

The skill is mostly automated:

1. Source-level prep (one-time per install) — extends ContainerConfig with env and blockedHosts fields and wires them into container-runner.ts so groups can pass per-group env overrides into Docker. Only runs if not already in place.
2. Per-group config — adds env and blockedHosts to groups/<folder>/container.json:

   {
     "env": {
       "ANTHROPIC_BASE_URL": "http://host.docker.internal:11434",
       "ANTHROPIC_API_KEY": "ollama",
       "NO_PROXY": "host.docker.internal",
       "no_proxy": "host.docker.internal"
     },
     "blockedHosts": ["api.anthropic.com"]
   }

3. Model selection — adds "model": "<your-model>" to the group’s shared Claude settings at data/v2-sessions/<group-id>/.claude-shared/settings.json. Claude Code reads its model from settings, not env vars.
4. Restart — the env block is read at container spawn, no rebuild needed.

What you get

• Local inference — the agent’s reasoning runs on your machine, never crosses the network.
• Per-group isolation — flip an env block to switch one group to local, leave others on Claude.
• Safety net — blockedHosts: ["api.anthropic.com"] resolves the host to 0.0.0.0 inside that group’s container so a misconfiguration can’t accidentally fall through to paid API.
• No provider code — Ollama’s Anthropic-API compatibility means this is just configuration, not new TypeScript.

Tips

• The skill changes chmod 755 to chmod 777 on /home/node in the container Dockerfile (and rebuilds) so any host UID can write there. Ollama’s host integration sometimes runs as a non-1000 UID inside the container.
• Cold first call: Ollama lazy-loads models. Watching curl -s http://localhost:11434/api/ps shows when the model is in memory.
• Models trained on Claude conversations sometimes claim to be Claude. Add a line to groups/<folder>/CLAUDE.md telling the model what it is to head this off.
• If the agent responds but Ollama’s /api/ps shows no activity, double-check both NO_PROXY and no_proxy (lowercase) — some HTTP libraries only honor one.
• Reverting is just deleting the env and blockedHosts keys from container.json and the model key from settings, then restarting. No rebuild needed.
• Pair with /add-ollama-tool if you also want Claude (running on the cloud) to call out to an Ollama model as an MCP tool — that’s a different use case.