ratel
The self-improvement layer for AI agents. Ratel reads your agent's traces, finds what fails, recommends fixes and runs evals on real sessions before you ship. S
What is ratel?
The self-improvement layer for AI agents. Ratel reads your agent's traces, finds what fails, recommends fixes and runs evals on real sessions before you ship. Self-hosted: one app, one Postgres. Includes the TypeScript and Python SDK
Overview of ratel
What it is
Ratel is a context engineering layer for AI agents. It selects only the tools and skills relevant to each turn, recovering accuracy lost to tool overload and reducing token usage. It does not require a vector database or infrastructure.
Key features
- Progressive Disclosure: indexes tools and skills into a catalog that the agent searches per turn, injecting only matching capabilities instead of loading everything up front.
- In-process BM25 and semantic retrieval: uses BM25 by default; semantic and hybrid ranking are opt-in per catalog or per call.
- Facts: constant grounding information (e.g., a shop's address) is registered as facts and re-injected only when not already fresh in the transcript.
- No vector DB required: retrieval is fast and deterministic without external infrastructure.
- SDKs for TypeScript and Python: provides
@ratel-ai/sdkandratel-aipackages. - Adapters for Vercel AI SDK and Mastra.
Use cases
Reducing token costs and improving accuracy in AI agent systems that suffer from tool overload. Suitable for local, open-source, and frontier model setups.
Getting started
TypeScript:bashpnpm add @ratel-ai/sdkPython:bashpip install ratel-aiSee the repository README for full usage examples.
License and pricing
The ratel-ai-core engine is licensed under Apache-2.0. Everything else (SDKs, telemetry helpers, examples) is MIT. The project is open source.
Based on the project's public README and repository metadata.