Skip to content

README.md

Latest commit

 

History

History
57 lines (45 loc) · 2.27 KB

README.md

File metadata and controls

57 lines (45 loc) · 2.27 KB

Mini SWE Agent (Python) — ACP Bridge

A minimal Agent Client Protocol (ACP) bridge that wraps mini-swe-agent so it can be run by Zed as an external agent over stdio.

Configure in Zed

Add an agent_servers entry to Zed’s settings.json. Point command to the Python interpreter that has both agent-client-protocol and mini-swe-agent installed, and args to this example script:

{
  "agent_servers": {
    "Mini SWE Agent (Python)": {
      "command": "/abs/path/to/python",
      "args": [
        "/abs/path/to/agent-client-protocol-python/examples/mini_swe_agent/agent.py"
      ],
      "env": {
        "MINI_SWE_MODEL": "openrouter/openai/gpt-4o-mini",
        "MINI_SWE_MODEL_KWARGS": "{\"api_base\":\"https://openrouter.ai/api/v1\"}",
        "OPENROUTER_API_KEY": "sk-or-..."
      }
    }
  }
}

Notes

  • If you install agent-client-protocol from PyPI, you do not need to set PYTHONPATH.
  • Using OpenRouter:
    • Set MINI_SWE_MODEL to a model supported by OpenRouter (e.g. openrouter/openai/gpt-4o-mini, openrouter/anthropic/claude-3.5-sonnet).
    • Set MINI_SWE_MODEL_KWARGS to a JSON containing api_base: { "api_base": "https://openrouter.ai/api/v1" }.
    • Set OPENROUTER_API_KEY to your API key.
  • Alternatively, you can use native OpenAI/Anthropic APIs. Set MINI_SWE_MODEL accordingly and provide the vendor-specific API key; MINI_SWE_MODEL_KWARGS is optional.

Requirements

Install mini-swe-agent (or at least its core deps) into the same environment:

pip install agent-client-protocol mini-swe-agent
# or: pip install litellm jinja2 tenacity

Then in Zed, open the Agents panel and select "Mini SWE Agent (Python)" to start a thread.

Behavior overview

  • User prompt handling
    • Text blocks are concatenated into a task and passed to mini-swe-agent.
  • Streaming updates
    • The agent sends session/update with agent_message_chunk for incremental messages.
  • Command execution visualization
    • Each bash execution is reported with a tool_call (start) and a tool_call_update (complete) including command and output (returncode in rawOutput).
  • Final result
    • A final agent_message_chunk is sent at the end of the turn with the submitted output.

Use Zed’s “open acp logs” command to inspect ACP traffic if needed.