Skip to content
ThermiteThermitev0.1.3

Thermite for solo builders

Independent developers and researchers have more ideas to test than hours to test them. An AI coding assistant helps while you work with it; Thermite continues when you stop. Its agents run experiments in parallel, a judge measures each one against the best version so far, and you review what was tried, what was accepted and why.

Use cases

Section titled “Use cases”

Each of these runs on one computer, most of them without a GPU.

Why it fits

Section titled “Why it fits”
  • It runs on the computer you have. Linux or macOS with Docker or Podman, a few CPU cores and a few gigabytes free. A GPU is optional. Installing takes about twenty minutes, or give your own AI agent one line and let it do the steps (Install on one machine).
  • Start free. The first campaign tutorial uses a free model, with no account and no key. Later, use a model server on your own computer, a Claude or ChatGPT plan you already pay for, or an API key (Choose a model).
  • No kit to write for a first try. The New campaign page asks what to improve, where the project is and how to run it, and writes a simple kit and contract from your answers (Your own project).
  • You set the spend. Nothing stops a campaign by default, because good ideas come after long dry spells. Set a stop time, or a limit on devices or spending, when you want one (Set a budget or a stop time).
  • Every result is verified. An agent never grades its own work. The judge builds each change, runs it, checks its output and times it on the same machine as the best version, and a win must hold up a second time before it counts (Why agents never judge).
  • Steer from anywhere. Watch on the dashboard over your private network, or let your own AI assistant watch and steer through MCP with a token of its own (Tailscale / WireGuard, Let an agent run Thermite over MCP).
  • Nothing is thrown away. Every attempt is kept with its code and the judge’s reasons, so an idea that lost tonight can still be read, and tried again, next week.

Get started

Section titled “Get started”
  1. Install on one machine, and watch the judge’s smoke check give a dozen known changes, including deliberate cheats, their expected verdicts.
  2. Run your first campaign on the sample project, with a free model, and follow the agents on Live.
  3. Start a campaign on your own project from New campaign, and set a stop time.
  4. Review the run’s page: its champion, what each lane tried and the judge’s reasons. Take the result as a patch, or let it continue.

What to know first

Section titled “What to know first”
  • The agents run code that a model wrote, on your computer, and under compose that code can reach your local network. Read Known limits before you connect a model.
  • With a paid API key, set a stop time and a limit at the provider first.
  • A Claude or ChatGPT plan may serve only its owner’s own Thermite.
See what a long campaign can doQwen3.8-27B on a $100 GPU: 131 changes judged in five days, on a Claude plan, a ChatGPT plan and a local model server, with no per-token bill.