Start here

Thermite lets AI agents work on a project for as long as it takes, while a judge they cannot touch decides which of their changes really count. This page helps you find where to begin.

What you need

• A computer running Linux or macOS, with a few CPU cores, a few gigabytes of free memory, and several gigabytes of free disk. Thermite’s three images take about 2 GB.
• Docker or Podman, which run Thermite’s containers. You download Thermite’s images ready-made, and build nothing.
• For the command line, from the second tutorial on: Node.js 22.19 or later, and git.
• About twenty minutes for the first run.

A GPU is optional, and so is a paid AI model: you can start with a free one.

Before you give Thermite a model, read Known limits. Agents run the code that a model writes on your computer, and under compose that code can reach your local network. The security model says what each part of Thermite can reach and what it trusts.

Pick your path

Try it on one computer

You want to see what Thermite does before you commit to anything. The four tutorials build on each other, in this order:

1. Your first run: start Thermite and watch its judge check a rig. Needs only the computer above.
2. Hand the work to agents: connect a free model, start a campaign on the sample project in Thermite’s repository, and watch agents propose changes. Needs tutorial 1 and an internet connection; no account or key.
3. Your first goal campaign: set a goal, submit a change yourself, and watch the campaign end when the goal is met. Needs tutorial 1; no model.
4. Your own project: a campaign on a small project of yours, started from the New campaign page. Needs tutorials 1 and 2, and a public repository that your computer can push to and Thermite can fetch over https.

What Thermite is gives the ideas behind it, in one minute.

Run it on your GPUs

You have machines with GPUs and want Thermite to use them.

1. Install Thermite on one machine, so it keeps running.
2. Let rigs on other machines reach control. This needs an HTTPS address for control, behind a reverse proxy you provide, and Thermite’s images in a registry that the rigs can pull from without a login. A release’s published images are such images, and its compose file gives them to control.
3. Add a rig on a machine over SSH or on a Kubernetes cluster, or run one on a RunPod GPU, paid by the hour.
4. Give a rig several GPUs, and read Sharing a device to choose how.

Note

Rigs that control sets up on machines and clusters, and rigs you join yourself, have no results repository yet. Their exports go to control instead (Get a rig’s results from control), but they can’t send each other champions. The rigs of the local compose stack have one. See goal 061.

Give it your own project

You have a project with a number you want to improve, or goals you want met.

1. First do tutorials 1 and 2 under Try it on one computer. Then Your own project: a first campaign from the New campaign page.
2. Write a kit for your project: how Thermite builds, runs and checks it.
3. Write a contract: what the campaign must achieve.
4. Why agents never judge: what makes a result count.

Let an agent run Thermite

You want your own AI assistant to watch and steer campaigns for you.

1. Let an agent run Thermite over MCP.
2. Tokens and roles: what an agent may and may not do. Only you approve what matters.
3. Point your agent at llms.txt, which lists every page of this site.

Words you will meet

Word	Meaning
Rig	Where the research happens: two containers on a machine, with the device the work is measured on.
Harness	The rig’s container where the judge and the agents work.
Supervisor	The rig’s other container: its link to control, which also does all its git work.
Control	The one service that shows and steers every rig: a dashboard, a REST API and MCP.
Owner’s token	The secret you sign in to the dashboard with. Only the owner approves what matters.
Campaign	One piece of research: what to improve or achieve, and how to judge it.
Contract	The file that sets out a campaign: what to improve or which goals to meet, which files agents may change, and its limits.
Run	A campaign running on one rig.
Kit	The judge’s recipe for a project: how to build it, run it, time it and check it. Agents never see it.
Lane	One line of search inside a run. Each lane’s agent works one trial after another.
Attempt	One change an agent submits.
Judge	The code that measures each change and gives it a verdict. Agents cannot change it.
Champion	The best confirmed result of a run so far. The starting point is champion #1.

The glossary has every term.