Every coding agent platform wants its machine described differently. Claude Code takes a session-start hook, Codex has its own settings, Ona uses tasks, Devin a blueprint. A team that runs two or three of them ends up keeping two or three setups in sync by hand, and they drift.
Preconfiguration is built for that. You describe the project once in preconfig.yaml and it writes each platform’s files. Then it proves the result works by running the project’s tests on a clean machine.
0.2, four more platforms
Preconfig 0.2 adds Claude Code, OpenAI Codex, Ona and Devin. The Claude Code output was run in a real cloud session, which turned up a detail no doc mentions: the image ships Python 3.13, so the setup script now handles that version. Codex, Ona and Devin outputs were checked against their published formats but haven’t run on those platforms yet, and the project says so plainly.
Old specs with no targets listed still produce the original five files, so nothing breaks on upgrade.
| Platform | What Preconfig writes | Tested how |
|---|---|---|
| Claude Code | Session-start hook for cloud sessions | Run in a real session |
| OpenAI Codex | Environment settings | Checked against documented format |
| Ona | Tasks | Checked against documented format |
| Devin | Blueprint | Checked against documented format |
0.3, a check on every pull request
Version 0.3 ships as a GitHub Action. On each pull request it checks the agent setup files and marks problems on the exact lines. When the spec or setup script changes, it runs the setup on a clean machine. If that fails, Preconfig Doctor’s diagnosis goes into the job summary, so you see why without digging through logs.
It also warns when its own knowledge of the platform formats is more than 60 days old. Agent platforms change their formats often, and a checker that quietly falls behind is worse than none.
The review caught a real risk before release: the verify step ran untrusted setup scripts with repository secrets under pull_request_target. That was fixed, along with fifteen other problems. A Windows line-endings bug, found by the action’s own test, went out in 0.3.1.
Why this market is open
Agent environments are young and fragmented. Each platform ships its own format and its own setup docs, and nobody owns the layer in between yet. If the platforms agree on one shared format, a translator matters less. Until then, one spec checked in CI beats four files edited by hand.
All 38 studio projects are listed on our Software Projects Desk.