Open-source sandbox for AI coding agents

Full agent autonomy,
safely contained.

Run Claude, Codex, OpenCode, and other coding agents at full autonomy. Each one runs in its own isolated container, with your files and network access under your control.

$ enclave starts an isolated container and drops you into your agent
Install Enclave View on GitHub

The problem

The freedom that makes agents useful makes them dangerous

An agent is most productive when it can act without asking. But the same freedom that makes it useful is what makes it dangerous.

Agents need broad access to be useful

To get things done end-to-end, an agent runs commands, installs packages, and edits files on its own. Gate every step behind a confirmation prompt and you lose most of the benefit.

Broad access on your host is dangerous

That same agent can delete the wrong files, leak a secret, be hijacked by a prompt injection, or collide with another agent sharing your machine.

The approach

Isolation that stays out of the agent's way

Enclave gives each session its own container, an isolated view of the filesystem, and its own network gateway. The agent keeps full autonomy inside; the damage it can do stays boxed in.

Isolated by the OS

Each session runs in its own container with a separate filesystem and network stack. Only the paths you mount, usually the project worktree, are visible to the agent.

Separate worktrees

Give each parallel agent its own git worktree, using plain git with no extra setup, so their changes never collide. You review and merge on your own terms.

Network you control

A sidecar gateway filters DNS and proxies outbound traffic. Agents reach only the domains on your allowlist; everything else is blocked and logged.

Sessions that persist

Start, pause, resume, and inspect a session whenever you want. Auth, history, and config persist across restarts, including custom skills and MCP servers.

Under the hood

How it works

Developer CLI manage CONTAINER AI coding agent Claude · Codex · OpenCode · … Mounted git worktree Default Docker backend Isolated filesystem and network all traffic GATEWAY DNS allowlist Transparent proxy Allowed Blocked Developer CLI manage CONTAINER AI coding agent Claude · Codex · OpenCode · … Mounted git worktree Default Docker backend Isolated filesystem and network all traffic GATEWAY DNS allowlist Transparent proxy Allowed Blocked

In the default Docker setup, each session runs in its own container behind a gateway that filters and logs its network traffic.

Quickstart

One command to a sandboxed agent

Point Enclave at any project. It builds the environment and starts your agent in a container on your current checkout.

# Start the default agent in an isolated container
enclave

# Pick a specific agent
enclave --tool codex

# Resume a previous session
enclave continue

# List active sessions
enclave ps

On the roadmap

Meet HomeShell

Today Enclave is a CLI. HomeShell is a graphical companion on the way: group your work into projects, split each into workstreams, and run agent sessions across them. Start, watch, and switch between sandboxed sessions with a live overview of every agent's status.

In development
HomeShell: a graphical dashboard listing Enclave projects, workstreams, and running agent sessions

Capabilities

Everything in the box

Supported agents

  • Claude Code (Anthropic)
  • Codex CLI (OpenAI)
  • OpenCode
  • Theia AI, Pi, and more
More

Command line

  • enclave starts the default agent
  • enclave --tool codex picks an agent
  • enclave continue resumes a session
  • enclave ps lists active sessions
More

Configuration

  • Custom skills and agent templates
  • Mount extra directories such as docs or related projects
  • Devcontainer support for existing setups
  • Per-tool and per-project overrides

Networking

  • Domain allowlists
  • Per-tool and per-project policies
  • Bridge host services into containers
  • DNS and request logging

Governance

Compliance and audit readiness

A foundation you can audit

Container isolation, network allowlists, and gateway logs already limit what a misaligned or injected agent can do, and they leave an evidence trail you can inspect: in restricted Docker sessions, DNS queries and proxied requests are logged, and access outside the allowlist is blocked. That same foundation can grow toward requirements like the EU Cyber Resilience Act (SBOM evidence and vulnerability reporting) and the EU AI Act (Article 12 record-keeping and Article 15 robustness against unauthorized use), with goals such as:

  • Log every outbound network request
  • Track downloaded dependencies and packages
  • Monitor git pull and push operations
  • Keep a complete audit trail of agent activity
  • Provide evidence for SBOM and supply-chain reviews
  • Govern the skills, MCP servers, and plugins available per session

Give your agents room to work, but safely.

Install Eclipse Enclave and start your first sandboxed session.