← All posts

I run a lot of AI coding agents. On a good day there are half a dozen of them going at once — one refactoring a module, one chasing a flaky test, one spiking a feature I’m not sure about yet. Each lives on its own git worktree, which is the easy part. The hard part is everything around the code.

Because the moment two agents try to boot the same app, they collide. Same 5432 for Postgres. Same 3000 for the web server. Same Redis, same volumes, same half-migrated database. One agent’s docker compose up steps on another’s, a test suite connects to the wrong database, and I lose ten minutes untangling which .env belongs to which branch. Multiply that by five agents and the workflow around the agents becomes the bottleneck — not the agents themselves.

I kept solving it by hand: bespoke port numbers per branch, a scratch .env here, a docker compose -p prefix there, a tmux tab I was afraid to close. It worked, barely, and it didn’t scale past a few agents. So I built the tool I wanted instead.

Meet Berth

Berth gives every agent its own berth — an isolated dev environment it can dock in without bumping into any other. For each parallel task you get:

• an isolated git worktree on a fresh branch,
• an isolated dev environment with auto-assigned, non-colliding ports,
• isolated state — its own container network, database, cache, and volumes,
• one command up, one command down, and
• a live view of what’s running where.

In practice it looks like this:

$ berth up agent-1

Ports:
  BACKEND_PORT    60291
  FRONTEND_PORT   60292

URLs:
  http://localhost:60291
  http://localhost:60292

✓ Berth 'agent-1' is running.

Spin up agent-2 and it gets its own worktree, its own free ports, and its own database — no collisions, nothing to configure. berth ls shows every berth, its branch, status, age, and ports. berth down agent-1 tears the whole thing down — services, volumes, worktree, branch — in one command. Abandoned an agent’s work? berth down --force discards it cleanly.

Under the hood it’s deliberately thin: git worktree for branches, Docker Compose for the environment, a port allocator, and a tiny state file. It reads a small berth.yml that points at your existing compose file and names the ports to inject. That’s the whole thing.

What it isn’t

Berth is not an agent runner, and it’s not a cloud platform. It’s the environment layer that sits underneath whatever agents you already use — Claude Code, Codex, Cursor, Aider — on purpose. I don’t want to replace your workflow; I want to stop it from tripping over itself. Staying tool-agnostic is the entire point.

Free, open source, built in the open

Berth is MIT licensed and free, and it’ll stay that way. It’s a single local binary — no daemon to babysit, no account, no bill. I built it because I needed it, I use it every day, and the honest best outcome is that it’s useful to other people drowning in the same parallel-agent mess. If that’s you, the code is on GitHub — star it, try it on your own projects, and tell me what breaks. Issues and PRs welcome.

This is the second thing we’ve shipped at OutpostLabs, after Resurface. Different surface — one lives in Jira, one lives in your terminal — but the same idea behind both: a small, focused tool that does one real thing well and gets out of your way.

More soon. If you want to follow along as I build these in the open, the blog is the place.