Ratel: retrieve your agent's tools each turn, cut ~80% of input tokens
The project topping Hacker News today swaps full tool catalogs for BM25 retrieval — no vector DB, runs as an MCP server, with one licensing catch.
Copy markdown
Retrieve, don't stuff
Ratel sits between your agent and its tool catalog and returns only the top-K relevant tools each turn via a deterministic BM25 index, so a 100+ tool setup collapses to roughly 2 tools per turn. You call `search_tools(query)` mid-turn or pre-filter the tools at the start of a message.
The numbers it reports
On the MetaTool benchmark the project reports a local Qwen model jumping from 8% to 77% tool-selection accuracy on a 100-tool catalog, and frontier models cutting input tokens ~82% and cost ~68% with only a minor accuracy tradeoff. Figures are self-reported, so measure against your own catalog.
No vector DB, no service
Retrieval runs entirely in-process: a Rust core (`ratel-ai-core`) does the BM25 indexing, wrapped by a TypeScript SDK — no embeddings, no vector database, and no extra network hop on the retrieval path.
Drop in as an MCP server
`@ratel-ai/mcp-server` exposes your catalog as an MCP server that works with Claude Code, Cursor and ChatGPT, so you can front an existing agent without rewriting it — `pnpm add @ratel-ai/sdk` to start, or `cargo add ratel-ai-core` for Rust.
The licensing catch
Ratel ships under Elastic License 2.0 with an open-source exemption: free for OSI-approved open-source projects, but commercial use needs a paid license — worth checking before you build it into a product.