Graph-based tool retrieval for LLM agents, reducing token usage by 79% with 82% accuracy.
graph-tool-call is worth checking the docs before setup with trust notes worth reviewing. Check agent compatibility and use-case fit before adding it to your workflow.
gh repo view SonAIengine/graph-tool-call --webOpen the official README and confirm the supported install method.
Add the server entry to your MCP client config.
Restart your agent and verify that the server tools appear.
Repository setup guidance
Looks usable, but maintenance, license, or security notes deserve a closer look.
Last commit was about 24 days ago.
16 GitHub stars indicate community interest.
0 open issues signal maintenance load.
Official MCP reference servers from Anthropic. Includes servers for filesystem, GitHub, Postgres, Slack, and more.
A secure MCP server for AI applications to interact with MySQL databases.
Up-to-date code documentation for LLMs and AI code editors, eliminating outdated or hallucinated API references.
This tool helps AI agents find the right functions to call. Instead of searching for tools one by one, it builds a graph of how tools connect, so it can find not just the tool you need but also the tools that provide the required inputs. It also shrinks the tool descriptions to save tokens, making the AI faster and cheaper.
graph-tool-call is a Python library that solves the problem of retrieving relevant tools from large catalogs for LLM agents. It treats tool retrieval as a graph problem, not just flat similarity search. It ingests tools from OpenAPI, GraphQL, MCP, Python functions, and structured catalogs, then builds a graph of dependencies between tools. When a user asks for something, it finds the target tool and also the prerequisite tools that produce its required inputs. It uses hybrid signals (keyword, graph, optional embeddings, MCP annotations) for retrieval, and then projects schemas to fit the model's token budget. It provides adapters for OpenAI, Anthropic, LangChain v1, MCP, Docker, and Kubernetes. The core package has zero dependencies, and optional integrations are available. It includes a demo that runs without any model or API key.
Graph-based tool retrieval for LLM agents, reducing token usage by 79% with 82% accuracy.
LLM application developers building agents with many tools., Platform engineers integrating tool catalogs with OpenAI or Anthropic APIs., Developers using LangChain or MCP who need efficient tool retrieval.
The setup section provides repository-level starting guidance, not a guarantee of an independently verified installation. Check the official README and release notes.
The library itself does not handle authentication or authorization; the host application must enforce security policies. Ingesting untrusted OpenAPI or MCP schemas could introduce malicious content; validate sources before use.
Similar or complementary options to evaluate include MCP Server Collection, mysql_mcp_server, context7.
Aug 20, 2026
MIT license detected.
3 security/trust notes recorded.
Setup difficulty is 3/5.
16
Stars
4
Forks
0
Issues
MIT
License