# Ratel AI

Ratel AI is the context-engineering platform for AI agents. On every turn it loads only the right tools, skills, and memory — and routes to the right model — so agents stay lean, accurate, and cheaper to run. Open-source (MIT), with measured input-token cuts of 70–85%. Advanced analytics and managed solutions available in the cloud offering.

Context engineering is the practice of deciding what belongs in a model's context window on each turn. As an agent accumulates tools, skills, and memory, its context bloats — accuracy drops, token bills rise, and behaviour becomes hard to debug (the failure mode known as "context rot"). Ratel injects only what's needed, when it's needed, using deterministic BM25 lexical retrieval that runs in your own process, alternatively using vector or hybrid retrieval as well.

Ratel works across four surfaces: (1) tool retrieval — surface only the relevant tools per turn across APIs, MCPs, Agents, Databases, Functions; (2) skill retrieval — load only the skills that matter each turn; (3) memory — rank prior context (decisions, preferences, artifacts) into the turn, a context layer Ratel orchestrates; (4) model routing — hint the right model and effort per request. It ships an open-source Rust core with TypeScript (`@ratel-ai/sdk`) and Python (`ratel-ai`) SDKs. Measured results (ratel-bench): 70–85% input-token reduction across models; −79% tokens and −64% cost on claude-sonnet-4-6 at pool=180; single-tool selection accuracy on a 100-tool catalog rising from 8.3% to 76.7% (qwen3.5, local).

## Documentation
- Introduction: What Ratel does and how it fits into an agent stack.
- Quickstart: Install and wire Ratel into an agent in a few minutes.
- Skills: The open-source skills suite — audit an agent codebase, plan observability, roll Ratel out across Claude Code, Cursor, and Codex.
- Python SDK: `ratel-ai` API reference for Python agent stacks.
- TypeScript SDK: `@ratel-ai/sdk` API reference for Node and edge agents.
- Full documentation (Markdown): Every docs page concatenated into one file — built for AI coding assistants (Cursor, Claude Code, Copilot) to ingest in a single fetch.

## Product
- Ratel AI homepage: What Ratel is — context rot, the four surfaces, benchmarks, and use cases.

## Benchmarks & proof
- Ratel benchmarks: Interactive token-cost and accuracy results across models.
- ratel-bench repository: Benchmark methodology and the full data behind the headline numbers.

## Source code
- Open-source core (GitHub): The Rust core and SDKs you run in-process.
- Skills suite (GitHub): The open-source skills that assess agents and roll Ratel out, used via coding agents.
- Ratel AI on GitHub: All of Ratel's open-source projects.

## Company & community
- X (Twitter): Product news and updates.
- LinkedIn: Company profile and announcements.
- Talk to the team: Book a call — for teams running agents in production.

## Optional
- Terms of Service: Legal terms.
- Privacy Policy: Privacy policy.