Ollama

Ollama runs open-source LLMs locally on your machine. It exposes an OpenAI-compatible API at localhost:11434, which AgentXchain connects to via the api_proxy adapter.

Which adapter?

api_proxy with provider: "ollama" — AgentXchain sends governed turn prompts to Ollama's local OpenAI-compatible endpoint.

Prerequisites

• Ollama installed (brew install ollama or download from ollama.com)
• A model pulled (ollama pull qwen3:32b or your preferred coding model)
• Ollama server running (ollama serve)
• agentxchain CLI installed

Configuration

{
  "runtimes": {
    "ollama-dev": {
      "type": "api_proxy",
      "provider": "ollama",
      "model": "qwen3:32b",
      "auth_env": "OLLAMA_API_KEY"
    }
  },
  "roles": {
    "dev": {
      "runtime": "ollama-dev",
      "mandate": "Implement features and fix bugs",
      "authority": "proposed"
    }
  }
}

Auth note

Ollama doesn't require an API key by default. Set a dummy value:

export OLLAMA_API_KEY="ollama"

The auth_env field is required by the adapter contract, but Ollama ignores the value.

Recommended coding models

Model	Size	Best for
qwen3-coder:32b	32B	Best local coding model
deepseek-coder-v3:33b	33B	Strong code generation
codestral:22b	22B	Fast Mistral coding model
llama4-scout:17b	17B	Meta's efficient coding model

Verify the connection

ollama serve &  # ensure server is running
export OLLAMA_API_KEY="ollama"
agentxchain connector check

Minimal working example

ollama pull qwen3:32b
ollama serve &
export OLLAMA_API_KEY="ollama"

mkdir my-project && cd my-project
agentxchain init --governed --template cli-tool --goal "Build a file renamer" -y
git init && git add -A && git commit -m "initial scaffold"
agentxchain run

Custom endpoint

If Ollama runs on a different host or port:

{
  "runtimes": {
    "ollama-remote": {
      "type": "api_proxy",
      "provider": "ollama",
      "model": "qwen3:32b",
      "auth_env": "OLLAMA_API_KEY",
      "base_url": "http://192.168.1.100:11434/v1/chat/completions"
    }
  }
}

Gotchas

• Model size vs. quality: Larger models produce better governed turn results but are slower. For QA/review roles, smaller models (7B-14B) may suffice. For implementation roles, use 32B+ models.
• Context window: Most Ollama models have 4K-32K context windows. AgentXchain dispatch bundles can be large. Check your model's context limit and set budget.max_tokens_per_turn accordingly.
• No internet required: Ollama runs entirely locally. This is ideal for air-gapped environments or privacy-sensitive codebases.
• GPU memory: Ensure you have enough VRAM for your chosen model. 32B models typically need 20GB+ VRAM.