Skip to content

Isolated, ready-to-run git worktrees for coding agents#

One directory per worktree name — cut fresh from origin, dependencies warm from a shared cache, your agent launched inside. Host-native, or sealed in a Docker sandbox.

Install See the full cycle

A small tree growing inside a glass box
treebox lifecycle demo: create launching Claude Code, enter with Codex, ls, and an interactive teardown

Name the branch up front (treebox create fix-auth) or don't: an unnamed worktree gets its own directory and a treebox/<petname> placeholder branch that can't be pushed — the agent renames it when the work takes shape. Agents work the same repo in parallel without collisions — on a laptop or over plain SSH.

Why treebox#

  • One directory per worktree

    git worktree under the hood — parallel agents, no shared files, indexes, or dev servers. Branches are named later, when the work has taken shape.

  • Never silently stale

    create branches from a fresh origin/<base> after a required fetch; a failed fetch exits loudly instead of building on old refs.

  • Warm in seconds

    Installs hardlink from shared caches (~/.cache/uv, the pnpm store, …). enter re-syncs only when the lockfile changed (an interrupted setup is finished rather than skipped).

  • The sandbox config lives outside the box

    The container is rendered from your operator template, never mounted — a boxed agent can't edit its own cage.

Two isolation modes, one pipeline#

Provisioning is identical either way; a pluggable isolation mode decides where the agent runs:

--isolation Sandbox Agent runs in
host (default) none the worktree shell
docker sandboxed a docker container, with your .env + caches mounted 
treebox create fix-auth                    # branch fix-auth off fresh origin/main
treebox create fix-auth --isolation docker   # same provisioning, sandboxed

Docker isolation needs exactly one extra thing: Docker — see what each isolation mode needs.

What it is — and isn't#

treebox does

  • Provision isolated worktrees, one per worktree name
  • Branch from a fresh origin/<base> — never stale refs
  • Install deps from shared caches (uv, npm, pnpm, go, cargo)
  • Copy your .env and submodules into place
  • Launch claude / codex with your subscription login
  • Speak script: stable exit codes, --json, --dry-run

treebox doesn't

  • Manage API keys — agents launch with your subscription login
  • Review, merge, or push your branches
  • Trust the target repo's config — its container config and hooks are ignored
  • Isolate anything in host isolation — that's what docker is for
  • Replace CI, or orchestrate fleets of agents
  • Install a package manager behind your back

Sixty seconds, end to end#

treebox doctor                       # verify the host is ready
treebox create                       # provision + launch claude (name generated)
treebox enter brave-otter --harness codex   # re-enter later; -H overrides the recorded harness for this session
treebox list                         # what exists, what's stale
treebox teardown brave-otter --delete-branch

Five commands cover the core lifecycle — with template for docker sandbox templates and version alongside. Install it, or read how it works.

Built by Seth Peters — an operator-focused layer for coding-agent infrastructure: git worktrees, sandbox boundaries, subscription auth, and repeatable dev environments. Building in this space? Connect on LinkedIn.