• 2 min read
Ratel aims to cut agent tool bloat at inference time
Ratel indexes tools and skills so agents only load what each turn needs, reducing token use and avoiding accuracy drops from overloaded prompts.
Image: Hacker News
Ratel is pitching a simple fix for a common agent problem: too many tools stuffed into every prompt. The open-source project says agents often pay to carry tool schemas, skills, and instructions they never use, which drives up token costs and can hurt accuracy as context windows fill up.
Instead of front-loading everything, Ratel builds a catalog of tools and skills and progressively reveals only the capabilities relevant to a given turn. According to the project, that approach can reduce token usage and recover accuracy lost to “tool overload,” and it does so without a vector database or extra infrastructure. Benchmark results are published at benchmark.ratel.sh.
The system separates tools from skills. When an agent needs to act, it calls search_capabilities. Ratel then searches dedicated indexes for tools and skills and returns a narrower set of matches. Tools can be invoked directly by ID, while skill instructions remain out of context until the agent explicitly loads them with get_skill_content.
By default, indexing uses BM25, the search ranking algorithm widely used in traditional search engines. Ratel applies it to schema-aware tool metadata along with skill names, descriptions, and tags. The project says this keeps retrieval fast and deterministic. Semantic and hybrid ranking are available as opt-in options, either per catalog or per call, using dense indexes backed by an in-process model or an OpenAI-compatible embedding endpoint.
Recommended reading
Hassabis says STEM makes you 10x better at AI
The project ships SDKs for TypeScript and Python, with a Rust core retrieval engine underneath. The repository is organized into:
- core/ for the Rust engine
- sdk/ts/ for the @ratel-ai/sdk package
- sdk/python/ for the ratel-ai package
- telemetry/ for OpenTelemetry conventions and helpers
There are also related projects: ratel-local, described as a local distribution for coding agents in front of an MCP setup, and ratel-bench, the benchmark harness behind the public benchmark site.
For developers building locally, the listed prerequisites are Rust stable, Node 24+, pnpm 10.28+, and for the Python SDK, Python 3.9+ and uv. Licensing is split: ratel-ai-core is under Apache-2.0, while the SDKs, telemetry helpers, and examples are released under MIT.
AI Editor
Ava covers the rapidly evolving world of artificial intelligence, from foundational models and research labs to the real-world economics of intelligence. With a background in computational linguistics, she cuts through the hype to find out what actually works. She firmly believes that benchmarks are just marketing until reproduced in the wild.
via Hacker News
