Running parallel AI coding agents sounds like a productivity win until you actually try it. The moment two Claude Code instances both reach for port 3000, one of them dies. The other inherits an .env file that isn’t there because you gitignored it. And the third agent, the one trying to run migrations, just discovered there is no database yet. Git worktrees isolate your code history fine. They do nothing about your runtime environment.
Sculptor, released October 1 by Imbue, takes a different approach: one Docker container per agent, full stop. Each agent gets its own copy of your repository, its own isolated runtime, its own installed dependencies, and its own ports. No collisions, no mysterious environment failures. When you want to bring an agent’s work back into your local repo, you click once and Pairing Mode syncs it via Mutagen. That’s the whole pitch.
Why Containers and Not Worktrees #
The gap between “worktree isolation” and “real isolation” matters more than it sounds. A git worktree gives each agent a separate working directory and branch. But they still share your machine’s file system, running services, and network stack. Spin up five worktrees running Next.js dev servers and you’ve got five processes each defaulting to port 3000. The first one wins; the rest throw errors.
It gets worse. Worktrees don’t carry over node_modules. They don’t get your .env. Every new worktree means reinstalling dependencies and recreating secrets by hand. For AI agents expected to run autonomously while you do other things, this is a documented failure mode, not an edge case.
Sculptor containers sidestep all of this. Each container gets a complete copy of the repository plus independent dependency installation. Port 3000 inside container A and port 3000 inside container B are different ports. No agent can see another agent’s database. The isolation is real.
Pairing Mode: The Feature That Makes It Useful #
Isolation that keeps agents separated forever would just be annoying. The reason Sculptor works in practice is Pairing Mode, which uses Mutagen to keep a container and your local repo synchronized in both directions. Enable it for a specific agent and changes flow back and forth in real time. You can open the agent’s files in your IDE, make edits, and the agent sees them. The agent commits something; your local directory reflects it.
This makes the workflow concrete. You open Sculptor, queue five tickets across five agents, then go do something else. Come back, review what each agent built, pair with the one closest to done, test it in your environment, merge it. Repeat. A beta tester put it plainly: “I was cooking dinner and blasting out features, coming back to see what Sculptor had cooked up for me.”
Getting Started #
Sculptor is free and open source under MIT. You need Docker v27 or newer and an existing Claude or Claude Max subscription — Sculptor doesn’t resell API tokens, it calls Anthropic directly with your credentials. Download the Mac (Apple Silicon) or Linux build from the Imbue site. Windows support exists via WSL but is not polished.
Claude Code support covers Opus 4.5 through 4.8. Codex was added in December 2025. A JavaScript or TypeScript plugin system lets you wire in other providers. The first container takes roughly a minute to build; subsequent ones are faster once the base image is cached. Three workflow skills come out of the box: spec writing, mock generation, and TDD bug fixing. A Suggestions feature (beta, on by default) runs an AI review pass on agent output before you merge.
What You Are Trading For Clean Isolation #
Docker isolation is not free. Containers consume more disk space and memory than worktrees, and that first per-container build minute adds up at scale. Managing multiple Postgres instances locally for agents that each need a database is your problem to solve — Sculptor does not handle database-per-agent provisioning out of the box.
Native GPT and Gemini support is not there yet. Claude is the first-class citizen; everything else requires custom harness configuration. Linux/Wayland users reported crashes. The Hacker News discussion reflects a tool in early but genuine use — real enthusiasm alongside real edge cases that have not been smoothed out.
What This Shift Actually Means #
The real unlock is not running more agents. It is changing what developers do while agents run. With proper isolation and real-time sync, you stop being a person who writes code and become a reviewer who evaluates what five autonomous agents produced in the last thirty minutes. That role shift is more significant than any individual Sculptor feature. The question parallel agentic tooling is quietly answering is: what does developer work look like when the bottleneck moves from writing to reviewing?
Sculptor is an early and imperfect answer to that question. But it is pointed at the right problem.