benchturn

For people who build open source with coding agents

We pay open-source engineers for the work they already do.

Your traces become RL environments in real time. You get a share of every sale.

Install benchturn Read the docs
curl -fsSL https://benchturn.com/install.sh | sh

Keys and .env files are never recorded. Any repo with a .gen-ignore file is skipped.

macOS or Linux. Takes about a minute. Read the script.

One command, about a minute

Install once, then keep working as you do now.

The script downloads one self-contained binary, signs you in with GitHub in the browser, and connects every coding agent it finds on the machine.

There are no tasks to pick up and nothing to annotate. Uploads run in the background, and your agent never slows down.

How it works

The best environments come from work that already happened.

AI labs train and test coding agents on environments: a real repository at a real commit, a task someone actually asked for, and a way to check the result. benchturn records your sessions on open-source projects, turns the good ones into environments, licenses them to labs, and pays you a share.

  1. 01

    Work with your agent

    Use Claude Code, Codex, Cursor or another supported agent on open-source code, the way you already do.

  2. 02

    Each turn is recorded

    The repository's tree before your prompt and after the agent stops, plus the transcript, so the change can be replayed exactly.

  3. 03

    A lab licenses it, you get paid

    Good sessions become environments offered to labs. When one is licensed, your share is added to your next payout.

What is recorded

Only what's needed to rebuild the task. Nothing more.

Work outside a git repository is not recorded, and a repository with a .gen-ignore file at its root is skipped entirely. benchturn ignore writes one for you.

Recorded, for each prompt

  • The repository's tree before the prompt and after the agent stops.
  • The transcript of that turn: your prompt, the agent's messages, and the tools it called.
  • The model name and a timestamp. At the end of a session, the full transcript.

Never recorded or shared

  • Keys, tokens, passwords and .env files. A secret scan runs on your machine before upload and again at ingest. A capture that trips it is dropped.
  • Your name. A lab receives the public repository, the prompts and the transcript, with no account information attached.
  • Anything in a repository you excluded. Private repositories can be recorded for you, but only public ones become environments.

Supported agents

Connects to the agents you already use.

benchturn init finds every supported agent on the machine and connects it. Safe to re-run; it finishes whatever isn't done.

  • Claude Code
  • Codex
  • Cursor
  • Gemini CLI
  • GitHub Copilot CLI
  • Cline
  • Pi
  • OpenCode

Two need one manual step, which benchturn init prints: Codex asks you to approve the hooks once with /hooks, and Cline has a Hooks switch under Settings → Features.

Where a session goes

Every session is reviewed.

We judge whether a session can become an environment: a real project, real code changed, a task that can be checked. We email you when a state changes.

  1. Received

    Uploaded and on your portal.

  2. Pending review

    We're judging whether it can become an environment.

  3. Environment built

    Rebuilt as a reproducible task.

  4. Listed

    Offered to labs.

  5. Licensed

    A lab bought it. Your share is added to your pending payout.

A session that is not a fit is deleted. We don't promise a date.

Questions

Things people ask first.

The full answers, and every command, are in the docs.

Do sessions from before I installed benchturn count?

Yes. Run benchturn past and your own agent reads the sessions already on disk from every supported product, picks the ones on public repositories with real work in them, checks each for private material, and shows you the list. Nothing is uploaded until you say yes to each one.

The instructions the agent follows are public at benchturn.com/agents.md.

What makes a session a good fit?

A real project that other people use or could use, with tests or some way to check a change, that builds and runs on Linux. And real work on it: a bug fixed, a feature added, or a refactor carried through. Best of all, you steered: corrected the agent, rejected its approach, or knew something it did not.

Will it slow my agent down?

No. Uploads run in the background on a timer. Nothing waits on them.

Can I stop recording, or delete something?

benchturn pause and benchturn resume stop and restart recording. benchturn delete <id> removes one capture from your machine and tells us to forget it. benchturn uninstall removes the hooks from every agent and leaves your own agents' files exactly as they were.

Where do captures live?

On your machine under ~/.claude/gen until they are uploaded, and a copy stays there afterwards.

Your best work is already on disk.

Install benchturn and keep working. Sessions upload on their own.

curl -fsSL https://benchturn.com/install.sh | sh