ratel
Context engineering for AI agents. ~80% fewer tokens. Fix tool overload. Skills and memory with in-process BM25 and semantic retrieval. Progressive Disclosure. No vector DB.
What is it?
What it is
A context engineering layer for AI agents that selects only relevant tools and skills for each turn.
Why it exists
To reduce costs by minimizing token usage from unnecessary tool schemas and instructions, and to improve accuracy by preventing model drift caused by tool overload.
Who should use it
AI agent developersLLM application engineersDevelopers building complex tool-calling agentsEngineers looking to reduce LLM token costsDevelopers experiencing model drift due to large context windows
Who should avoid it
Developers with very simple agents that only use 1-2 toolsUsers who require a managed vector database solution
How it works
A quick walkthrough in plain English
How ratel works
Step 1 of 3
You interact with it
Open ratel, send a request, or connect it to your stack.
Features
Advantages
- Open source (MIT)
- Active Rust ecosystem
- Self-hosted deployment options
Disadvantages
- Requires operational ownership for self-hosted setups
- Community support varies by project maturity
Installation
native
TypeScript: pnpm add @ratel-ai/sdk Python: pip install ratel-ai
FAQ
How does Ratel reduce costs for AI agents?
Ratel selectively discloses only the tools and skills relevant to each turn, avoiding the cost of sending all tool schemas, skills, and instructions upfront on every call.
What retrieval method does Ratel use for tool/skill selection?
Ratel uses the BM25 algorithm, a search engine-grade method, to efficiently and deterministically retrieve relevant tools and skills based on metadata and descriptions.
How does Ratel improve model accuracy?
By filtering out irrelevant tools and skills per turn, Ratel prevents context bloat and model drift, ensuring the agent focuses on task-critical capabilities.
Which programming languages does the Ratel SDK support?
The SDK supports TypeScript and Python, with examples and documentation provided for both in the repository.
Is a vector database required for Ratel's functionality?
No, Ratel operates without a vector database, relying on BM25 for retrieval and avoiding the overhead of vector indexing.
How are skill instructions incorporated into the agent's context?
Skill instructions are loaded dynamically via the `get_skill_content` tool only when relevant to a specific turn, keeping context lean and task-focused.
Featured in Videos
YouTube tutorials and walkthroughs for ratel
Alternatives
Similar projects ranked by category, topics, and text overlap.
