Foundry: how one founder runs an engineering team of Claude Code agents

I run the engineering team for my SaaS company. The team is me and a handful of Claude Code agents.

The hard part was never getting agents to write code. It was knowing what's in flight, what's verified, and what's safe to ship. So I built a system to manage them, and I've open-sourced it.

Why "Foundry"

It's where issues get cast into shipped code. It's also how one founder runs a whole engineering team.

The flow

Foundry turns a GitHub issue into a verified pull request:

1. I write an issue and comment "@foundry go"
2. Foundry starts Claude Code in an isolated git worktree
3. The agent works and files a typed report: done, needs input, plan ready, or blocked
4. Foundry runs lint, typecheck, build, and tests. No PR until it's green.
5. It opens the PR, stands up a preview, and waits for me
6. I approve, or comment and it goes back to work

Merging stays a human decision. Always.

Running it from my phone

Because everything happens on GitHub, I run it from my phone: one comment per issue, and a verified PR and a live preview waiting when I look. Foundry keeps a status comment up to date with the agent, branch, PR, preview, run cost and a timeline of each run.

Setup

It's a GitHub Action, so there's no server to run. It runs Claude Code on your own subscription, and it's MIT licensed. I'm rolling it out across Joynt's TypeScript platform and our Unity project.

Try it

• Repo: https://github.com/joyntshops/foundry
• Demo: https://github.com/joyntshops/foundry-demo (the agent-built PRs still have their runs and previews attached)

Issues, feedback, and stars all welcome.