Your AI agent In a Box

Run AI agents in secure, isolated containers on any codebase to get work done.

openbox — zsh
$ openbox run --repo ./my-app --agent claude-code \       --prompt "Add rate limiting to /api/users"   ◆ Building image from agents/claude-code/Dockerfile... ◆ Copying workspace to /tmp/openbox-a3f9c2/ ✓ Box ready — box_a3f9c2   ● claude-code I'll add rate limiting using express-rate-limit...   ⚙ write_file  src/middleware/rateLimit.ts   ⚙ edit_file   src/routes/users.ts   ⚙ run_command npm install express-rate-limit   ✓ Done $0.018 · 3 files changed · 12 turns ? Sync changes to ./my-app? y
Works with
claude-code
codex
gemini
opencode
custom agents
Features

Isolated by default.
Powerful by design.

OpenBox manages containers, streams results, and tracks every change — so you just write the prompt.

Agent-agnostic

Bring any AI agent — Claude Code, Codex, Gemini, or your own. Just add a Dockerfile and an agent.yaml. No vendor lock-in.

Persistent boxes

Containers persist after the run. Browse the workspace, inspect diffs, resume where you left off, or cherry-pick changes on your terms.

Interactive REPL

Have a multi-turn conversation with your agent inside a live container. No cold starts, no context loss between prompts.

Context injection

Feed your agent OpenAPI specs, docs, or any reference file with --context. Bundled into the workspace so the agent can use them immediately.

Streaming SDK

Embed OpenBox in any Node.js app. Stream tool calls, token counts, and cost in real time. Stateless and dependency-injected — safe for concurrent server use.

API source sync

Point at a Postman collection or OpenAPI spec. When the API changes, openbox sync regenerates your client code automatically.

Web UI

A built-in Next.js dashboard with live chat, diff viewer, and file preview. Self-host it or run it locally.

How it works

From prompt to diff
in seconds

01

Box is created

OpenBox copies your local repo (or clones a GitHub URL) to an isolated temp workspace, then builds a Docker image from the agent's Dockerfile.

02

Agent runs inside

The container mounts the workspace, receives your prompt and context files, and runs the agent. Structured JSON events stream back in real time.

03

Inspect the diff

After the agent finishes, view a full git diff. The box persists — resume the session, run another prompt, or explore the workspace freely.

04

Sync on your terms

When satisfied, sync changes back to your source directory or open a GitHub PR. Clean up when you're ready.

Documentation

Everything you need
to get started

Full CLI reference, SDK API docs, and agent guides — all in one place.

Ready to open the box?

Install in seconds. Your first agent run takes under a minute.

$ npm install -g openbox
View on GitHub