Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.nanoclaw.dev/llms.txt

Use this file to discover all available pages before exploring further.

In v2, Ollama moved from a core integration to a skill on the providers branch. Install via /add-ollama-provider (run Ollama as a full provider for an agent group) or /add-ollama-tool (keep Claude as the agent, expose Ollama as a tool). The MCP setup described below reflects v1 mechanics; a v2 rewrite is pending. See Integrations overview for the v2 provider model.
NanoClaw can delegate tasks to local models running on Ollama, while Claude remains the orchestrator. This lets you offload cheaper tasks (summarization, translation, general queries) to local models and reduce API costs.

​
How it works

The /add-ollama-tool skill adds a stdio-based MCP server inside the agent container. The MCP server exposes these tools:
ToolDescriptionRequires
ollama_list_modelsLists all locally installed Ollama models—
ollama_generateSends a prompt to a specified model and returns the response—
ollama_pull_modelPulls (downloads) a model from the Ollama registryOLLAMA_ADMIN_TOOLS=true
ollama_delete_modelDeletes a locally installed model to free disk spaceOLLAMA_ADMIN_TOOLS=true
ollama_show_modelShows model details: parameters, template, architecture infoOLLAMA_ADMIN_TOOLS=true
ollama_list_runningLists models currently loaded in memory with resource usageOLLAMA_ADMIN_TOOLS=true
The container agent reaches Ollama on the host via host.docker.internal:11434. Claude decides when to use local models based on task complexity — you don’t need to configure routing rules.

​
Prerequisites

Install Ollama and pull at least one model: 
# Install Ollama (macOS/Linux)
curl -fsSL https://ollama.com/install.sh | sh

# Pull a model
ollama pull llama3.2        # Good general purpose (2GB)
ollama pull gemma3:1b       # Small and fast (1GB)
ollama pull qwen3-coder:30b # Best for code tasks (18GB)
Verify Ollama is running: 
ollama list

​
Installation

Apply the skill in Claude Code: 
/add-ollama-tool
Or manually: 
git fetch upstream skill/ollama-tool
git merge upstream/skill/ollama-tool
This adds:
  • container/agent-runner/src/ollama-mcp-stdio.ts — MCP server that bridges to Ollama
  • scripts/ollama-watch.sh — macOS notification watcher for Ollama status
  • Ollama MCP configuration and [OLLAMA] log surfacing in the agent runner
  • OLLAMA_ADMIN_TOOLS config option for model management tools
After merging, rebuild the container: 
./container/build.sh

​
Configuration

Set OLLAMA_HOST in .env if Ollama runs on a non-default address: 
# Default (usually correct — no need to set)
OLLAMA_HOST=http://host.docker.internal:11434
The MCP server automatically falls back to localhost if host.docker.internal fails. To enable admin tools (pull, delete, show, list running), add: 
OLLAMA_ADMIN_TOOLS=true
Ollama must be running on the host before starting NanoClaw. The MCP server writes status to /workspace/ipc/ollama_status.json so the host process can surface connection issues in logs.

​
Usage

Once installed, Claude can use local models transparently. For example:
“Summarize this document using a local model”
Claude will call ollama_list_models to see available models, then ollama_generate with the appropriate prompt. You can also be explicit about which model to use:
“Use llama3.2 to translate this to Spanish”

​
Third-party model endpoints

Independently of Ollama, NanoClaw supports any Anthropic API-compatible endpoint. Set these in .env: 
ANTHROPIC_BASE_URL=https://your-api-endpoint.com
ANTHROPIC_AUTH_TOKEN=your-token-here
This allows you to use:
  • Open-source models on Together AI, Fireworks, etc.
  • Custom model deployments with Anthropic-compatible APIs
When using custom endpoints, the secret injection layer (OneCLI Agent Vault in v1.2.35+, or the credential proxy in earlier versions) still intercepts container API requests. Ensure the endpoint is reachable from the host.
Last modified on April 23, 2026