acp-mcp

<div align="left"> <h1>ACP to MCP Adapter</h1> **Connect ACP Agents to MCP Applications Seamlessly** [](https://github.com/i-am-bee/beeai-framework?tab=Apache-2.0-1-ov-file#readme) [](https://bsky.app/profile/beeaiagents.bsky.social) [](https://discord.com/invite/NradeA6ZNF) [](https://lfaidata.foundation/projects/) </div> The **ACP to MCP Adapter** is a lightweight standalone server that acts as a bridge between two AI ecosystems: **Agent Communication Protocol (ACP)** for agent-to-agent communication and **Model Context Protocol (MCP)** for connecting AI models to external tools. It allows MCP applications (like **Claude Desktop**) to discover and interact with ACP agents as resources. ## Capabilities & Tradeoffs This adapter enables interoperability between ACP and MCP with specific benefits and tradeoffs: ### Benefits - Makes ACP agents discoverable as MCP resources - Exposes ACP agent runs as MCP tools - Bridges two ecosystems with minimal configuration ### Current Limitations - ACP agents become MCP tools instead of conversational peers - No streaming of incremental updates - No shared memory across servers - Basic content translation between formats without support for complex data structures This adapter is best for situations where you need ACP agents in MCP environments and accept these compromises. ## Requirements - Python 3.11 or higher - Installed Python packages: `acp-sdk`, `mcp` - An ACP server running (Tip: Follow the [ACP quickstart](https://github.com/i-am-bee/acp/blob/main/README.md#quickstart) to start one easily) - An MCP client application (We use [Claude Desktop](https://claude.ai/download) in the quickstart) ## Quickstart **1. Run the Adapter** Start the adapter and connect it to your ACP server: ```sh uvx acp-mcp http://localhost:8000 ``` > [!NOTE] > Replace `http://localhost:8000` with your ACP server URL if different. <details> <summary><strong>Prefer Docker?</strong></summary> ```sh docker run -i --rm ghcr.io/i-am-bee/acp-mcp http://host.docker.internal:8000 ``` **Tip:** `host.docker.internal` allows Docker containers to reach services running on the host (adjust if needed for your setup). </details> **2. Connect via Claude Desktop** To connect via Claude Desktop, follow these steps: 1. Open the Claude menu on your computer and navigate to Settings (note: this is separate from the in-app Claude account settings). 2. Navigate to Developer > Edit Config 3. The config file will be created here: - macOS: `~/Library/Application Support/Claude/claude_desktop_config.json` - Windows: `%APPDATA%\Claude\claude_desktop_config.json` 4. Edit the file with the following: ```json { "mcpServers": { "acp-local": { "command": "uvx", "args": ["acp-mcp", "http://localhost:8000"] } } } ``` <details> <summary><strong>Prefer Docker?</strong></summary> ```json { "mcpServers": { "acp-docker": { "command": "docker", "args": [ "run", "-i", "--rm", "ghcr.io/i-am-bee/acp-mcp", "http://host.docker.internal:8000" ] } } } ``` </details> **3. Restart Claude Desktop and Invoke Your ACP Agent** After restarting, invoke your ACP agent with: ``` use "echo" agent with the "Good morning!" input ``` Accept the integration and observe the agent running.   > [!TIP] > ACP agents are also registered as **MCP resources** in Claude Desktop.<br /> > To attach them manually, click the Resources icon (two plugs connecting) in the sidebar, labeled "Attach from MCP", then select an agent like `acp://agents/echo`. ## How It Works 1. The adapter connects to your ACP server. 2. It automatically discovers all registered ACP agents. 3. Each ACP agent is registered in MCP as a resource using the URI: `acp://agents/{agent_name}` 4. The adapter provides a new MCP tool called `run_agent`, letting MCP apps easily invoke ACP agents. ## Supported Transports - Currently supports Stdio transport --- Developed by contributors to the BeeAI project, this initiative is part of the [Linux Foundation AI & Data program](https://lfaidata.foundation/projects/). Its development follows open, collaborative, and community-driven practices.