Native ACP Agent Reference Implementation

crow-cli

Minimal. Extensible. Native. A reference implementation of the Agent Client Protocol with zero frameworks—just pure Python, OpenAI SDK, FastMCP, and SQLAlchemy.

Documentation →
Or run directly with uvx crow-cli

Get Started in Seconds

Installation
$ uv tool install crow-cli --python 3.14
# Or run without installing:
$ uvx crow-cli
Initialize Project
$ crow init
# Creates project structure with:
# - Docker Compose with SearXNG
# - Agent configuration
# - Minimal boilerplate

Built for Simplicity

No bloat. No unnecessary abstractions. Just the essentials for building native ACP agents.

SearXNG Included

Built-in web browser capabilities via Docker Compose. No API keys required for search.

FastMCP

All tools are build on top of FastMCP because you don't need a tool calling framework in 2026.

OpenAI SDK

Direct integration with OpenAI's Python SDK. No wrapper frameworks between you and the API.

SQLAlchemy

Robust database management with SQLAlchemy ORM. Store agent state, memory, and logs efficiently.

Docker Ready

Production-ready Docker Compose setup out of the box. Isolate your agent environment completely.

Extensible Design

Minimal core, maximum flexibility. Reference architecture that's easy to extend and customize.

Architecture Philosophy

Zero Framework Dependencies

We believe in transparency. crow-cli uses only openai, acp, fastmcp, and sqlalchemy. No black-box frameworks hiding the implementation details.

Native ACP Implementation

Full compliance with the Agent Client Protocol specification. Works with any ACP-compatible client.

SearXNG Integration

Unlimited search via self-hosted SearXNG instance in Docker. API search keys? We don't need no stinking API search keys.

Coming Soon

The Ecosystem is Growing

Spec Kit Integration

Standardized agent specification tooling

ACP Registry

Discover and share agent configurations