openclaw-claude-code

An OpenClaw plugin that integrates Claude Code into your messaging workflow. Control Claude Code from Slack (or any OpenClaw-supported channel) to analyze, plan, and execute coding tasks on your local machine.

Features

• claude_plan — Read-only analysis mode. Analyze codebases, review architecture, plan implementations. No files are modified.
• claude_exec — Execution mode. Implement features, fix bugs, refactor code. Operates with full permissions within whitelisted directories.
• claude_teams — Multi-agent mode. Multiple Claude Code agents collaborate in parallel with built-in file locking and coordination.

Security Model

All security constraints are enforced in code — the LLM cannot bypass them.

Layer	What it controls	Default
allowedPaths	Directories where Claude Code can operate	[] (deny all — must configure explicitly)
maxTimeoutSecs	Hard cap on task duration	600 (10 min)
maxConcurrent	Max simultaneous Claude Code processes	2
Path validation	Resolves ~ and blocks ../../ traversal	Always on

Permission Modes

Tool	Claude Code mode	Behavior
claude_plan	plan	Read-only. Cannot create, edit, or delete files.
claude_exec	bypassPermissions	Full write access within allowedPaths directories only.
claude_teams	bypassPermissions	Same as exec, with multi-agent coordination.

Prerequisites

• OpenClaw installed and running
• Claude Code installed locally (claude --version)
• Node.js 22+

Installation

git clone https://github.com/Phoenizard/openclaw-claude-code.git
cd openclaw-claude-code
npm install
openclaw plugins install -l .
openclaw gateway restart

Verify:

openclaw plugins list
# Should show: claude-code: loaded

Configuration

Edit ~/.openclaw/openclaw.json:

{
  "plugins": {
    "entries": {
      "claude-code": {
        "enabled": true,
        "config": {
          "allowedPaths": ["~/projects/my-app", "~/projects/another-app"],
          "maxTimeoutSecs": 600,
          "maxConcurrent": 2,
          "claudeOauthToken": "your-token-here"
        }
      }
    }
  },
  "agents": {
    "list": [
      {
        "id": "main",
        "default": true,
        "tools": {
          "allow": ["claude_plan", "claude_exec", "claude_teams"]
        }
      }
    ]
  }
}

Authentication

Claude Code needs authentication to call the Anthropic API. Since the Gateway runs as a system service (launchd/systemd), it doesn't inherit your shell environment. Set claudeOauthToken in the plugin config, or set CLAUDE_BIN to point to a pre-authenticated Claude Code binary.

Claude Binary Resolution

The plugin searches for the claude binary in this order:

1. $CLAUDE_BIN environment variable
2. ~/bin/claude (recommended — create a copy here to avoid spaces in path)
3. ~/Library/Application Support/Claude/claude-code/<latest>/claude
4. claude on PATH

macOS note: The Claude Desktop install path contains a space (Application Support). Node.js spawn() may fail with ENOENT. Copy the binary to ~/bin/claude to avoid this issue.

Usage

Send messages to your OpenClaw bot in Slack:

Analyze (read-only)

Use claude_plan to analyze ~/projects/my-app architecture

Execute a task

Use claude_exec to add error handling to ~/projects/my-app/src/api.ts

Multi-agent collaboration

Use claude_teams in ~/projects/my-app: Frontend team builds the login page. Backend team implements the auth API.

Plan-then-Execute workflow (recommended)

1. Plan: Ask claude_plan to create an implementation plan
2. Review: Read the plan in Slack, request changes if needed
3. Execute: Ask claude_exec to implement the approved plan

Tool Reference

Tool	Mode	Default Timeout	Use Case
claude_plan	plan (read-only)	300s	Architecture analysis, code review, planning
claude_exec	bypassPermissions	300s	Feature implementation, bug fixes, refactoring
claude_teams	bypassPermissions	600s	Parallel development (frontend + backend + tests)

All tools accept:

• task (required) — Task description
• workdir (optional) — Working directory (must be in allowedPaths)
• timeout (optional) — Timeout in seconds (clamped to maxTimeoutSecs)

License

MIT