Repository navigation
FAQ
Is it just another MCP server? It is a gateway: it sits between your agents and your APIs (and other MCP servers), exposes them through one endpoint, and learns from how agents use them. The new part is turning repeated multi-call workflows into single, verified tools.
Does it use an LLM? No, not by default. Import, discovery, evaluation and capability matching are deterministic. An optional LLM stage (off by default) can read free-text API docs and help match capability requests.
Will it change my data on its own? No. Write and destructive tools stay hidden until a human grants them, every destructive call needs its own approval, and the evaluator never runs a plan that writes.
What does it store?
Tool names, timings, redacted arguments (or none, with
captureArguments: "none") and keyed fingerprints of results, so it can
see that one call's output fed another. It never stores tool results.
Traces are deleted after 14 days by default.
How many repetitions before it finds a workflow?
Three sessions by default (patterns.minSupport), with the steps linked by
data flow.
Do the speedups include model time? No. The benchmarks have no LLM in the loop, so they understate real agents: each avoided call is also an avoided model turn.
Does it work with local models? Yes. Toolsmith doesn't care which model the client uses. Fewer tool calls help most where each turn is slow or error-prone.
Which clients work? Any MCP client over stdio or Streamable HTTP, on protocol 2025-11-25 or 2026-07-28. See Connect Your Client.
Can I run it for a team? Yes: per-agent tokens or OAuth 2.1, per-client allow-lists and rate limits, and PostgreSQL for several nodes. See Team Deployment.
Is it production-ready? It is tested (unit, integration, end-to-end and security suites, on SQLite and PostgreSQL) and was validated over fifteen independent review rounds, but it has not yet been hardened by production use.
How does it compare to other tools? See COMPETITIVE_ANALYSIS.md and the research it builds on in PAPERS.md.
How can I help? Try it on your API and open an issue with what worked and what didn't. Stars help others find it.
toolsmith-mcp · MIT · Repository · Report an issue · Changelog
Getting started
Using it
Operating it
Links