Cut down on repeated AI model calls for an agent that performs the same kind of task often.
Turn recorded AI agent behavior into a small, fast, verified program.
Run AI-agent-derived logic safely inside a restricted sandbox environment.
Track and cap real dollar spend on AI model calls used as a fallback path.
| rightnow-ai/auto | yarlabs/hyperspace-db | xqnode/codex-helper | |
|---|---|---|---|
| Stars | 116 | 113 | 110 |
| Language | Rust | Rust | Rust |
| Setup difficulty | hard | hard | easy |
| Complexity | 4/5 | 5/5 | 2/5 |
| Audience | developer | developer | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires a pinned Rust toolchain and an exact-version flatc compiler, paid fallback paths need an OpenAI API key.
Auto is a tool that watches an AI agent work and turns the repeated, predictable parts of what it does into small, fast, cheap programs, so the expensive AI model does not have to be called again for the same kind of task. It first records what an agent actually does when performing a task: which tools it calls, what arguments it passes, what results come back. It then checks which of those actions always turn out the same way given the same input, since those parts do not really need an AI model at all, just ordinary code. Those predictable parts get extracted into small, verified programs, and anything left over that still varies gets compressed into much smaller specialist models. The result is a compact package, called a cbin, containing code, small models, and a report describing what was measured about it. When Auto runs, it first tries the compiled, fast path for a given input. If that input looks unfamiliar or too different from what was recorded before, it falls back to the original, larger AI model, records what happens this time, and folds that new information back into the compiled version so the same situation is not solved from scratch again next time. The project calls this a ratchet: nothing gets figured out twice. The compiled programs run inside a sandbox, a restricted environment that can only do what it is explicitly allowed to do, so a compiled program cannot exceed its stated capabilities. According to the measurements reported in the project's own results, the cost of running each task fell from about 59 millionths of a dollar using the large AI model every time, down to about 2 millionths of a dollar per task once the compiled path took over, while matching the reference AI model's answers on familiar inputs the large majority of the time. Response speed on the compiled path was also reported as dramatically faster than calling the AI model directly. This project is built with the Rust programming language and is aimed at people building AI agent systems who want to cut down on repeated AI model costs. The project's own documentation includes a research paper describing the measurements in detail. It is released under the Apache 2.0 license, which permits free use, including commercial use.
A tool that watches an AI agent work, converts its repeated predictable behavior into small fast programs, and only calls the expensive AI model again when it sees something genuinely new.
Mainly Rust. The stack also includes Rust, WebAssembly, Python.
Use freely for any purpose, including commercial use, as long as you keep the copyright notice.
Setup difficulty is rated hard, with roughly 1h+ to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.