OpenClaw Plugin: Claude Code

Run Claude Code in secure, isolated containers with your Claude Max subscription.

An OpenClaw plugin that executes Claude Code CLI sessions in rootless Podman containers. Let your AI agents delegate complex coding tasks to Claude Code without risking your host system.

Table of Contents

• Why Use This Plugin?
• FAQ
• Features
• Requirements
• Installation
• Configuration
• Authentication
• Job Completion Notifications
• Registered Tools
• Security
• Development
• Container Image
• Releasing
• Troubleshooting

Why Use This Plugin?

Use Your Claude Max Credits

Have a Claude Max subscription? This plugin lets you use your existing subscription credits for agentic coding tasks instead of paying per-token API costs. Your OAuth credentials are securely passed to containerized Claude Code sessions.

Contain the Blast Radius

Claude Code with --dangerously-skip-permissions can modify any file and run any command. This plugin contains each session in a rootless container with:

• All Linux capabilities dropped
• Configurable network isolation
• Memory and CPU limits
• Optional AppArmor profiles

Bad code, infinite loops, or accidental rm -rf stays inside the container. Your host system remains untouched.

Parallel Workspaces

Run multiple Claude Code sessions simultaneously, each in its own isolated workspace. Sessions maintain state across interactions, enabling complex multi-step workflows without interference.

Works With Any OpenClaw Agent

Any OpenClaw agent can use the claude_code_start tool to offload coding tasks. The orchestrating agent stays lightweight while Claude Code handles the heavy lifting in its own container.

FAQ

Why Podman instead of Docker?

Security. This plugin runs AI-generated code with --dangerously-skip-permissions — the container is your only safety net. Podman is rootless by default: no daemon, no root process, no privilege escalation path. If an AI agent escapes the container, it lands in an unprivileged user namespace with no capabilities.

Docker's default mode runs a root daemon. A container escape from a root-daemon Docker setup gives the attacker full root access to the host. Docker can run rootless, but it requires additional setup and isn't the default. Most Docker forks of this plugin skip that step.

The plugin supports runtime: "docker" for users who have configured rootless Docker, but Podman is strongly recommended.

Why use this plugin instead of the coding-agent skill?

The built-in coding-agent skill is a prompt that teaches an OpenClaw agent to delegate coding tasks using the platform's existing bash and process tools. It's lightweight, supports multiple agents (Codex, Claude Code, Pi), and requires zero setup beyond loading the skill.

This plugin solves a different problem: containment. When Claude Code runs with --dangerously-skip-permissions, it can modify any file and run any command. The coding-agent skill runs those agents directly on your host (or in OpenClaw's sandbox), while this plugin runs each session in a rootless Podman container with all capabilities dropped, resource limits enforced, and /tmp mounted as tmpfs.

Use the coding-agent skill when:

• You want multi-agent support (Codex, Claude Code, Pi)
• Quick setup matters more than isolation
• You're already comfortable with OpenClaw's sandbox mode
• Tasks are short-lived and don't need session persistence

Use this plugin when:

• You're running Claude Code with --dangerously-skip-permissions and want real containment
• You need persistent sessions that survive across multiple interactions
• You want structured job management (status, output pagination, activity detection, crash recovery)
• You're running untrusted or experimental code and need hard resource limits

They can also work together: an agent could use the coding-agent skill for quick Codex tasks while routing longer Claude Code sessions through this plugin for isolation.

Features

• Isolated Execution: Each Claude Code session runs in its own container with dropped capabilities
• Real-time Streaming: Output is captured as it's generated using Claude Code's stream-json format
• Session Persistence: Sessions maintain state across multiple interactions
• Dual Authentication: Supports both API key and OAuth/Claude Max credentials
• Resource Limits: Configurable memory, CPU, and PID limits
• AppArmor Support: Optional AppArmor profile for additional security hardening
• Automatic Cleanup: Idle sessions are automatically cleaned up

Requirements

• OpenClaw >= 2025.1.0
• Podman (recommended) or Docker
• Node.js >= 22

Installation

From npm

openclaw plugins install @13rac1/openclaw-plugin-claude-code

From GitHub Release

Download the latest release zip from GitHub Releases and extract to your plugins directory:

unzip openclaw-plugin-claude-code-*.zip -d ~/.openclaw/plugins/openclaw-plugin-claude-code

Container Image

The plugin requires a container image with Claude Code CLI installed. Pull the pre-built image:

podman pull ghcr.io/13rac1/openclaw-claude-code:latest

Or build it yourself:

podman build -t ghcr.io/13rac1/openclaw-claude-code:latest .

Configuration

Add to your openclaw.json:

{
  "plugins": {
    "enabled": true,
    "load": {
      "paths": ["path/to/dist/index.js"]
    },
    "entries": {
      "claude-code": {
        "enabled": true,
        "config": {
          "image": "ghcr.io/13rac1/openclaw-claude-code:latest",
          "runtime": "podman",
          "startupTimeout": 30,
          "idleTimeout": 120,
          "memory": "512m",
          "cpus": "1.0",
          "network": "bridge",
          "sessionsDir": "~/.openclaw/claude-sessions",
          "workspacesDir": "~/.openclaw/workspaces",
          "sessionIdleTimeout": 3600
        }
      }
    }
  }
}

Configuration Options

Authentication

The plugin supports two authentication methods:

1. OAuth / Claude Max (Recommended)

If you have Claude Max or enterprise OAuth credentials, place your credentials file at:

~/.claude/.credentials.json

The plugin mounts your ~/.claude directory directly into each container session, so OAuth token refreshes persist automatically.

2. API Key

Set the ANTHROPIC_API_KEY environment variable:

export ANTHROPIC_API_KEY=sk-ant-...

Note: If both are available, OAuth credentials take precedence.

Job Completion Notifications

The plugin sends webhook notifications when jobs complete. No polling required.

Setup

Enable webhooks in OpenClaw config (~/.openclaw/openclaw.json):

{
  "hooks": {
    "enabled": true,
    "token": "your-secret-token"
  }
}

The plugin automatically reads the token from the OpenClaw config file. When a job completes, fails, or is cancelled, the agent receives a message with the job status, duration, and output size.

Sandbox Tool Allowlist

If you use OpenClaw's sandbox mode (sandbox.mode: "non-main" or "all"), webhook-triggered agents run in a sandboxed context with restricted tool access. To allow the agent to use this plugin's tools, add them to your sandbox tool allowlist in ~/.openclaw/openclaw.json:

{
  "tools": {
    "sandbox": {
      "tools": {
        "allow": [
          "exec", "process", "read", "write", "edit", "apply_patch", "image",
          "sessions_list", "sessions_history", "sessions_send", "sessions_spawn", "session_status",
          "claude_code_start", "claude_code_status", "claude_code_output",
          "claude_code_cancel", "claude_code_cleanup", "claude_code_sessions"
        ]
      }
    }
  }
}

See Multi-Agent Sandbox & Tools for more details on sandbox tool policies.

Registered Tools

claude_code_start

Start a Claude Code task in the background. Returns immediately with a job ID.

Parameters:

• prompt (required): The task or prompt to send to Claude Code
• session_id (optional): Session ID to continue a previous session

Returns: { jobId: string, sessionId: string }

claude_code_status

Check the status of a running or completed job.

Parameters:

• job_id (required): Job ID returned from claude_code_start
• session_id (optional): Session ID

Returns:

• status: Job status (pending, running, completed, failed, cancelled)
• elapsedSeconds: Time since job started
• outputSize: Total output size in bytes
• tailOutput: Last ~500 chars of output (for quick preview)
• lastOutputSecondsAgo: Seconds since last output was produced
• activityState: "active" (producing output), "processing" (CPU busy), or "idle"
• metrics: CPU and memory usage
• exitCode: Process exit code (when completed)
• error: Error message (if failed)

claude_code_output

Read or tail output from a job.

Parameters:

• job_id (required): Job ID
• session_id (optional): Session ID
• offset (optional): Byte offset to start reading from (for tailing)
• limit (optional): Maximum bytes to read (default 64KB)

Returns: Output content with hasMore flag for pagination

claude_code_cancel

Cancel a running job and stop its container.

Parameters:

• job_id (required): Job ID
• session_id (optional): Session ID

claude_code_cleanup

Clean up idle sessions and their jobs.

claude_code_sessions

List all active sessions with age, last activity, message count, and active job info.

Security

The plugin implements multiple layers of security:

1. Rootless Containers: Uses Podman rootless mode by default
2. Capability Dropping: All Linux capabilities are dropped (--cap-drop ALL)
3. Resource Limits: Memory, CPU, and PID limits prevent resource exhaustion
4. tmpfs: /tmp is mounted as tmpfs with nosuid (512MB, exec allowed for compilers)
5. Network Isolation: Configurable network mode (can be set to none)
6. AppArmor: Optional AppArmor profile support for MAC enforcement

Development

Setup

git clone https://github.com/13rac1/openclaw-plugin-claude-code.git
cd openclaw-plugin-claude-code
npm install

Build

npm run build

Test

# Unit tests (mocked)
npm test

# Integration tests (requires Podman)
npm run test:integration

# All tests
npm run test:all

Local Development

Link the plugin for development:

openclaw plugins install -l ./path/to/openclaw-plugin-claude-code

Container Image

The included Dockerfile creates a Debian Bookworm-based image with:

• Node.js 22
• Claude Code CLI (npm global)
• Go 1.22.5 + TinyGo 0.32.0
• Python 3 with venv
• Common dev tools: git, ripgrep, jq, curl

Building Multi-arch Images

# Single architecture (current)
podman build -t ghcr.io/13rac1/openclaw-claude-code:latest .

# Multi-architecture (arm64 + amd64)
GITHUB_USERNAME=13rac1 ./scripts/build-and-push.sh --multi-arch

Releasing

Releases are automated via GitHub Actions when a version tag is pushed.

Prerequisites

1. Configure NPM_TOKEN secret in GitHub repository settings
2. Ensure you have push access to the repository

Release Process

1. Update version in package.json
2. Update CHANGELOG.md with release notes
3. Commit the changes:

   git add package.json CHANGELOG.md
   git commit -m "chore: release v1.x.x"
   

4. Create and push a version tag:

   git tag v1.x.x
   git push origin main --tags
   

The release workflow will automatically:

• Run tests
• Publish to npm with provenance
• Build and push multi-arch container images to ghcr.io

Troubleshooting

Container image not found

Error: Container image not found: ghcr.io/13rac1/openclaw-claude-code:latest

Pull or build the container image:

podman pull ghcr.io/13rac1/openclaw-claude-code:latest

No authentication available

Error: No authentication available. Set ANTHROPIC_API_KEY or have ~/.claude/.credentials.json

Either:

1. Set ANTHROPIC_API_KEY environment variable, or
2. Place OAuth credentials at ~/.claude/.credentials.json

Startup timeout

Error: startup_timeout - No output within 30 seconds

The container failed to start or produce output. Check:

• Container image is valid
• Sufficient system resources
• Network connectivity (if network: bridge)

Increase startupTimeout if needed.

Idle timeout

Error: idle_timeout - No output for 120 seconds

Claude Code stopped producing output. This may indicate:

• Task completed but output wasn't captured
• Claude Code is stuck
• Task requires more time (increase idleTimeout)

License

Apache-2.0 - see LICENSE

Contributing

See CONTRIBUTING.md