Projects / Thermite
Thermite
Put AI agents to work on your code. Keep only what's proven.
Open-source autoresearch. Agents propose. A judge they can't see or change measures every attempt and keeps only what it confirms.
The receipts
Measured, with a public source for every number.
7.74 → 30.06 tok/s
One campaign: 131 judged attempts, 28 kept. Its winner became tensorless's engine.
receipt ↗Why I built it
I need an autoresearch platform to improve engines on the P100, so I can make local AI more affordable. I also need auto-scientist systems: deep research, hypothesis verdicts and more.
Agents are really good at making the number go up. Not always by making the work better. So in Thermite, agents propose, and code they can't touch does the judging.
What it is
Three campaign modes: a hill climb that pushes one metric with guards, a goal campaign that runs until a list of checks passes, and a research campaign that answers a question from web sources.
Each agent works in its own lane on its own copy of the code, and a planner keeps the experiments distinct.
The judge holds the clock. It builds each change, measures it against the current best on the same hardware, checks that the output is still correct, and re-measures wins on fresh runs and hidden test cases.
It runs on a laptop, GPU machines, a Kubernetes cluster or a RunPod pod. Agents can use a free model, your own model server, an API key, or a Claude or ChatGPT subscription.
One dashboard watches every campaign live, with a REST API, MCP and a thermite CLI. Every attempt is logged with its verdict.
How it works
You need Linux or macOS, Docker or Podman, and about twenty minutes. No GPU, paid model or account is needed for the first tutorials.
Start with the first run at usethermite.com.
Honest caveats
- Thermite is young and changes every day.
- Agents run model-written code on your machine. Run it where that's acceptable.