MCP-Zero

MCP-Zero is an open-source research project accompanying the paper "MCP-Zero: Active Tool Discovery for Autonomous LLM Agents." The repository, owned by xfey, releases the code for similarity-based tool retrieval along with a dataset of MCP server tools. It is released under the MIT License.

What It Is

MCP-Zero is a method and code base for letting LLM agents proactively discover and select tools instead of being given a fixed tool list. The released code implements the retrieval capability described in the paper: a matcher for similarity matching, a reformatter for tool description JSON, a sampler for selecting target tools, prompt guides, and experiment scripts for APIBank and an MCP-tools needle test.

MCP-tools Dataset

The project provides a dataset drawn from the MCP official repository, covering 308 servers and 2,797 tools after filtering. Each server entry includes a name, a summary derived from its README, a description, and embeddings (text-embedding-3-large, 3072 dimensions) for the description and summary. Each tool entry includes its name, description, description embedding, and parameters when explicitly defined. The dataset is downloaded from Google Drive and placed at a specified path. Scripts under MCP-tools/build_data let users build a custom dataset from MCP server READMEs, including a vLLM deployment script for Qwen2.5-72B-Instruct and an extraction prompt.

Current Status

The authors state that the released code covers the retrieval work in the paper. They say other modules aimed at industry application, such as dynamic deployment of MCP servers and environment deployment for GAIA testing, are still to be developed. A Hugging Face link for the dataset is listed as coming soon.