Self-hosted workspaces for parallel coding agents.
Run agents from the same commit, keep their changes after workspace cleanup, and send each change to GitHub as a draft PR. Built for platform teams running code migrations and maintenance jobs with their own agents.
Install v0.1.21-alpha.1 · Watch the 36-second demo · Signed Linux packages
Alpha for test repositories. Setup is manual on Linux with PostgreSQL. The two-agent walkthrough uses gVisor for execution isolation.
- Connected. Use your existing GitHub CI, review and merge process. Each saved change records its starting revision and agent run, and you can export it to Git after cleanup. Agents that need Git commands can use a workspace with private Git metadata.
- Standalone (experimental). Start an empty native repository, publish changes under TreeHarbor review and check policies, and export all branches to Git.
Use your own agent controllers to claim tasks, download pinned source and submit changes over HTTPS. Controller guide · Rust CLI · Experimental MCP adapter.
Copy-on-write workspaces reuse prepared source and give each agent a writable layer. In a Linux ARM64 development benchmark with four workspaces and 8,195 files, reusing the imported Git pack cut first preparation from 1.70–1.73 s to 0.85–0.87 s after import. Git worktree took about 0.51 s on those first reuse runs. Repeat batches took 0.14–0.16 s with TreeHarbor and about 0.65 s with Git worktree. Method and raw data
Anthropic’s Scion inspired the separation of workspaces from source history. OpenAI’s Pando reinforced the value of keeping existing Git workflows. TreeHarbor starts with the handoff from agent tasks to your existing review process. Scion · Pando
Deployment layout · Status and release gates · Roadmap · Contributing
Source-available under Elastic License 2.0.