Thermite

Thermite is an open-source autoresearch system that runs on your own machines. You point AI agents at a number to push, a list of goals to meet, or a question to answer. A judge the agents cannot read or change measures every attempt and keeps only confirmed wins. The project is Apache-2.0 licensed, and the current release is v0.1.3.

What It Is

Thermite automates the loop of changing something, measuring it, keeping it if it is better, and trying the next idea. Agents only propose changes. A separate judge builds each change, runs it against the best version so far on the same hardware, checks that output is still correct, and decides. Wins become the new starting point, called the champion, and losing attempts are kept with the judge's reasons. The champion can be downloaded as a patch for your project.

Campaign modes

Each campaign is one of three modes. A hill climb pushes one metric, such as speed, size or accuracy, as far as it goes while guards hold everything else steady. A goal campaign works until a list of checks passes or the evidence shows it cannot. A research campaign reads sources for hours or days, and code verifies each quote word for word against the archived page; the result is a paper plus a corpus that later campaigns can search. Modes can be chained.

How results are verified

The judge times each change and the champion in alternating blocks on the same device and decides with statistics. It times the champion against itself once a day to measure machine noise. Wins are re-run on fresh runs and on test cases the agents never saw. On a rig with an auditor, unusually large gains get a review, and users can require manual approval of every new champion. The repository ships a judge smoke check of known changes, including cheating attempts, with expected verdicts.

Deployment model

There is no hosted service and no account. A small control service provides a dashboard, REST API and MCP server, while rigs do the research on a local computer, machines over SSH, a Kubernetes cluster, or a RunPod GPU. Agents run in Pi, OpenCode, Claude Code or Codex, and can use a free model, a local server such as Ollama or vLLM, an API key, or a Claude or ChatGPT plan. It requires Linux or macOS with Docker or Podman. The README warns that the project is under heavy development.

Case study

The Thermite site reports a study with tensorless in which four campaigns tuned an inference engine for Qwen3.8-27B on one Tesla P100, raising decode speed from 7.74 to 30.06 tokens per second across 131 judged changes.