ratel

Introduction

The context engineering layer for AI agents. Selects only the tools and skills relevant to each turn, recovering accuracy lost to tool overload and cutting what you pay per call. No vector DB, no infra.

Why

• Cost: Every tool schema sent to the model is tokens you pay for. Fewer tools in context means lower spend on every call.
• Accuracy: Models get worse as tool lists grow. Some drop from 77% to 8% accuracy just from having too many options.
• Ratel fixes both: by indexing your full catalog and injecting only the tools that match the current task, keeping the rest out of context entirely.

Across local, open-source, and frontier model setups, Ratel cuts token usage and recovers accuracy lost to tool overload — without embeddings or a vector DB. Full results: benchmark.ratel.sh

Install

Building an agent in TypeScript or Python? Add the SDK:

pnpm add @ratel-ai/sdk

pip install ratel-ai

import { ToolCatalog, searchCapabilitiesTool, invokeToolTool } from "@ratel-ai/sdk";

const catalog = new ToolCatalog();
catalog.register({
  id: "read_file",
  name: "read_file",
  description: "Read a file from local disk.",
  inputSchema: { properties: { path: { type: "string" } } },
  execute: async ({ path }) => ({ contents: await fs.readFile(path, "utf8") }),
});

const search = searchCapabilitiesTool(catalog);
const invoke = invokeToolTool(catalog);

from ratel_ai import ToolCatalog, ExecutableTool, search_capabilities_tool, invoke_tool_tool

catalog = ToolCatalog()
catalog.register(ExecutableTool(
  id="read_file",
  name="read_file",
  description="Read a file from local disk.",
  input_schema={"properties": {"path": {"type": "string"}}},
  execute=lambda args: {"contents": open(args["path"]).read()},
))

search = search_capabilities_tool(catalog)
invoke = invoke_tool_tool(catalog)

Examples: Vercel AI SDK · Pydantic AI

Using Claude Code, Cursor, or ChatGPT with MCP servers? Drop Ratel in front of your existing setup with no code changes:

npx -y @ratel-ai/mcp-server mcp import

Full docs: ratel-ai/ratel-mcp

How it works

When your agent needs to act, it calls search_capabilities. Ratel searches its internal index and returns only the most relevant tools. The model sees a short, focused list and picks correctly far more often.

The index uses BM25 by default, the same algorithm behind most search engines, applied to each tool's name and description. It is fast, deterministic, and adds no latency to your agent loop. Semantic and hybrid ranking are opt-in per catalog or per call, running a local embedding model in the same process.

Full docs

The Ratel project

Ratel scales from an in-process library to a managed service — one engine, one catalog contract, all the way up:

The hosted cloud is decided direction (ADR-0002), not yet public; a standalone server is deferred (ADR-0003).

Repo layout

src/
├── core/              # ratel-ai-core — Rust BM25 engine
├── sdk/ts/            # @ratel-ai/sdk — TypeScript SDK (NAPI-bound)
├── sdk/python/        # ratel-ai — Python SDK (PyO3-bound)
└── telemetry/         # OTel conventions + helper packages
protocol/              # catalog-source wire contract
examples/              # End-to-end SDK examples
docs/                  # ADRs

Build & test

Prerequisites: Rust stable, Node 24+, pnpm 10.28+. Python SDK: Python 3.9+ and uv.

cargo build --workspace && cargo test --workspace   # Rust
pnpm install && pnpm -r build && pnpm -r test       # TypeScript
# Python: see src/sdk/python/README.md

Contributing

• CONTRIBUTING.md
• AGENTS.md — for coding agents working in this repo

License

The ratel-ai-core engine is licensed under Apache-2.0 — an explicit patent grant for the engine others embed. Everything else (SDKs, telemetry helpers, examples) is MIT. See ADR-0009 for the rationale.