If you’ve run two Claude Code sessions on the same machine, you already know the pain. Agent A spins up a dev server on port 3000. Agent B tries the same port five seconds later. One crashes. You spend the next ten minutes manually editing .env files, changing port numbers, restarting Docker containers. Meanwhile, both agents are sitting there waiting, burning tokens on nothing.
Git worktrees solved the code isolation problem months ago. Every major coding agent — Claude Code, Codex, Cursor — now supports parallel worktrees out of the box. But here’s what nobody talks about: the code is isolated, the runtime is not. Two worktrees still share the same localhost, the same port space, the same database, the same Docker daemon. Change a migration in worktree A, and worktree B’s integration tests break because they’re hitting the same Postgres instance.
Coasts fixes this. One CLI command, and each worktree gets its own containerized runtime — dedicated ports, isolated network, independent services. No cloud. No rewrites of your existing Docker Compose setup. Just full runtime isolation on your local machine.
The Problem Is Bigger Than Port Conflicts
The worktree-sharing-localhost issue isn’t a minor annoyance. It’s a fundamental blocker for the way AI coding is heading.
Right now the workflow looks like this: you fire up three or four agents in parallel, each working on a different feature branch in its own worktree. Agent one is refactoring the auth module. Agent two is adding a new API endpoint. Agent three is writing integration tests. This is the dream scenario. Multiple agents, multiple branches, one afternoon.
But the moment any of these agents needs to run the app — spin up a dev server, execute tests against a database, check if the UI renders correctly — everything falls apart. Port 5432 is already taken. The Redis instance has stale data from another branch. The migration from agent one just dropped a table that agent three’s tests depend on.
Developers have been hacking around this with manual port offset scripts: base port plus worktree index times ten, update every .env file, pray nothing collides. It works until it doesn’t. And it definitely doesn’t scale to five or six parallel agents, which is where things are heading.
Tools like Container Use from Dagger and worktree-compose have tried to address parts of this problem. Container Use gives each agent a containerized sandbox, but it’s more focused on sandboxing the agent itself rather than providing a full dev stack. worktree-compose auto-generates isolated Docker Compose stacks per worktree, but it’s zero-config at the cost of flexibility — you can’t easily share services across instances when you want to.
Coasts takes a different approach. It doesn’t sandbox the agent. The agent still runs on your host machine, with full access to your tools and environment. What Coasts containerizes is the runtime — the services your code depends on. Think of it as: the agent stays free, the infrastructure gets isolated.
How Coasts Actually Works
The architecture is straightforward. A background daemon (coastd) communicates with a thin CLI client over a Unix socket. The daemon manages state in a local database, handles port allocation, and orchestrates Docker containers. There’s also a web UI built with React and Vite for monitoring everything visually.
You define a Coastfile at your repo root — it’s TOML, dead simple. Point it at your existing docker-compose.yml, map your ports, and you’re done. If you don’t use Docker Compose, Coasts can work with standalone Docker setups too.
The key commands tell the story. coast build compiles your environment once. coast run spins up an isolated instance tied to a specific worktree. coast checkout binds the canonical ports (the ones your browser and tools expect) to whichever instance you want to interact with directly. coast ls and coast ps show you what’s running.
What makes this interesting is the switching strategies. Not every service needs to restart when you switch between worktrees. Coasts lets you configure per-service behavior: none (keep running as-is), hot (swap the bind mount without restarting), restart, or rebuild. The creator claims this brought worktree switching time from around two minutes down to about eight seconds. That’s the difference between a tolerable workflow and a broken one.
The shared services concept is clever too. Your Postgres or Redis can run once on the host Docker daemon instead of being duplicated inside every coast. Each coast gets its own database within that shared Postgres instance, so data stays isolated without the overhead of running five separate database containers. About 200MB of overhead per additional coast — not nothing, but manageable on any modern dev machine.
Secrets management is worth noting. Host-side scripts extract secrets at build time, stored in SQLite rather than baked into images. They’re injectable as environment variables or file-system writes, and can be re-injected without rebuilding. macOS Keychain integration is built in.
What the Hacker News Crowd Thinks
The Show HN discussion is revealing. The first question everyone asks is “why not just use Docker directly?” The creator’s answer highlights eight specific gaps that plain Docker can’t fill: no control plane for host-side port management, no per-service switching strategies, no way to preserve your original Docker setup without modification, no shared service optimization, no built-in observability UI.
The predictable “goodbye Mac users” comment about Docker-in-Docker came up immediately. The creator pushed back — says it works fine on Mac through Docker Desktop, OrbStack, or Colima, with no noticeable latency from the virtualization layer. Given that most developers running parallel AI agents are on Macs, this matters.
One interesting thread dug into database state management. When multiple agents modify shared databases, how do you handle schema conflicts? The answer: it’s configurable. Integration tests get fully isolated databases. UI work can share services. The orchestration of what-gets-shared-and-what-doesn’t is left to the developer, which feels like the right call — being too opinionated here would break more setups than it fixes.
The project explicitly positions itself as not a sandboxing tool. It’s complementary to host-side sandboxing solutions. Your agent still runs on your machine with full access. Coasts handles the runtime layer only. This is a deliberate design choice that keeps the tool focused and avoids the scope creep that’s killed similar projects.
The Competitive Landscape Is Getting Crowded
Coasts isn’t the only player trying to solve agent isolation. Dagger’s Container Use launched last year, giving each agent a containerized sandbox with its own worktree. worktree-compose takes the zero-config approach, auto-generating isolated Docker Compose stacks. parallel-code from Johannes Jo lets you run Claude Code, Codex, and Gemini side by side in separate worktrees. And cloud-based solutions like Daytona, Niteshift, and Devin’s built-in environments sidestep the local problem entirely by moving everything to remote containers.
But Coasts occupies a specific niche: local-first, runtime-only, harness-agnostic. It doesn’t care whether you’re using Claude Code or Cursor or Codex or whatever ships next month. The only requirement is git worktrees. No cloud dependency, no vendor lock-in, no AI upsells. MIT licensed, 161 stars on GitHub, written primarily in Rust (74.8%) with a TypeScript UI layer.
The project is also a Y Combinator company, which suggests there’s a commercial angle coming — though the current tool is entirely free and open source.
At 345 commits and version 0.1.45 as of March 30, 2026, this is early-stage software. Seven core maintainers. The install is a one-liner: eval a curl script that sets up the daemon and CLI. Whether it matures into essential infrastructure or gets absorbed into the agents themselves (Claude Code already has built-in worktree support — full runtime isolation could be next) is the open question.
But right now, if you’re running multiple AI coding agents in parallel and tired of port conflicts and shared database nightmares, Coasts is the most complete local solution available. The fact that it layers onto your existing Docker setup instead of replacing it means the adoption cost is basically zero. And in infrastructure tooling, low adoption cost wins every time.
You Might Also Like
- 27k Github Stars in Weeks Learn Claude Code by Shareai lab Breaks Down ai Coding Agents Into 12 Lessons
- Claude Channels Scores 375 Points on Hacker News Anthropics Play to Replace Openclaw
- 27 Agents 109 Skills 88k Github Stars is Everything Claude Code Genius or Over Engineering
- Claude hud hit 5 3k Github Stars Because Developers Were Flying Blind With Claude Code
- Ensu got 328 Points on Hacker News the Privacy Crowd Wants ai That Never Phones Home
Top AI Product 
Leave a comment