MCP-Zero enables autonomous LLM agents to discover and utilize tools by proactively constructing toolchains. It addresses the challenge of agents needing to identify and integrate relevant tools for complex tasks, benefiting researchers and developers building sophisticated AI systems.

How It Works

MCP-Zero employs a similarity matching approach to discover tools. It leverages pre-computed embeddings (from text-embedding-3-large) for tool descriptions and server summaries to find the most relevant tools for a given task. This method allows for efficient retrieval and selection of appropriate tools from a large corpus.

Quick Start & Requirements

• Install: Code is available via git clone.
• Prerequisites: Python, text-embedding-3-large model for embeddings. The project includes scripts for building custom datasets using vLLM and Qwen2.5-72B-Instruct.
• Dataset: Pre-built dataset available via Google Drive (link in README).
• Links: arXiv Paper

Highlighted Details

• Implements retrieval capabilities for tool discovery.
• Includes a dataset of 308 servers and 2,797 tools with embeddings.
• Provides scripts for building custom datasets from MCP server READMEs.
• Demonstrates proactive toolchain construction for tasks like "Making a great meal."

Maintenance & Community

The project is actively being updated, with plans for industry applications and expansion. The README encourages users to "Leave a star🌟 to let me know you are staying updated."

Licensing & Compatibility

The README does not explicitly state a license. The code is open-sourced for research purposes.

Limitations & Caveats

The project is described as having code released for the paper, with other modules (dynamic deployment, GAIA test environment) still needing involvement for broader industry application.