Ratel: open-source library gives AI agents unlimited tools without bloating context
Ratel is an open-source library that lets an AI agent keep a full catalog of tools and skills while showing it only the few that matter for each turn, a technique called . Its creators, who previously helped SaaS companies build agents, kept running into the same wall as they added more tools and instructions: the growing context caused frequent and skyrocketing token bills.
They tried like dynamically loading tools and splitting agents into subagents before building Ratel to solve the problem directly. It supports both keyword and , running entirely in-process with no extra needed.
One production user reportedly cut by up to 81% in their first month without losing accuracy. The library is .
Key points
- Uses : keeps the full tool/skill catalog but surfaces only what's relevant per turn
- Supports both keyword and
- Runs in-process with no additional required
- One production user cut up to 81% without losing accuracy
- open-source library