ClaudeClaw: Run Persistent Slack Agents with OS-Level Sandboxing
What It Does
ClaudeClaw is a Claude Code plugin that transforms Claude into a persistent agent orchestrator. It listens to messaging channels (Slack, WhatsApp, Telegram, Discord, Gmail), routes messages to isolated Claude agents, and manages conversations, memory, scheduled tasks, and webhooks—all from a single Node.js process.
The system started as a fork of NanoClaw but was completely rebuilt as a Claude Code plugin with a pluggable extension system. The entire codebase (~8K lines of TypeScript) fits within Claude's context window, making it fully comprehensible and modifiable by Claude itself.
Why It Matters For Claude Code Users
Most Claude Code users interact with Claude through their terminal. ClaudeClaw extends this capability to team communication platforms where work actually happens. Instead of switching contexts between Slack and your terminal, you can deploy Claude agents that live in your Slack channels, ready to handle queries, run code, or trigger workflows.
The key architectural improvement over previous solutions is the native sandbox runtime using Anthropic's @anthropic-ai/sandbox-runtime. This provides OS-level process sandboxing (Seatbelt on macOS, bubblewrap on Linux) with sub-10ms cold starts—no Docker containers, no VMs, no 2-5 second delays per agent invocation.
Setup & Installation
Install ClaudeClaw as a Claude Code plugin:
# Clone the repository
git clone https://github.com/yourusername/claudeclaw.git
cd claudeclaw
# Install dependencies
npm install
# Link as a Claude Code plugin
claude --plugin-dir .
The plugin uses a clean separation between plugin entry (src/index.ts) and service entry (src/service.ts). The plugin loads via claude --plugin-dir and returns immediately—no side effects. The service runs the message loop as a persistent background process.
Key Features You Can Use Today
Pluggable Extension System: Extensions register capabilities without modifying core code. The triage system, webhook triggers, and cost tracking are all extensions—the orchestrator core doesn't know about them.
Structured Memory: Agents have more than just CLAUDE.md. They get search capabilities, daily logs, topic files, and memory tools that persist across conversations.
Per-Group Agent Configuration: Different Slack channels can use different Claude models, tools, and behaviors. Your #engineering channel might use Claude Opus 4.6 with full code execution, while #support uses Claude Sonnet 4.6 with limited tools.
Cost Tracking: Every agent run tracks token usage and estimated cost, giving you visibility into operational expenses.
Webhook Integration: Trigger agents from CI/CD pipelines, monitoring alerts, or any webhook source. This bridges the gap between your automation systems and team communication.
When To Use It
- Team Support: Deploy a Claude agent to your team's Slack channel that can answer technical questions, run code snippets, or check documentation.
- CI/CD Notifications: Have Claude analyze build failures or deployment logs and provide actionable insights directly in Slack.
- Scheduled Tasks: Run daily standup summaries, weekly report generation, or regular database checks.
- Multi-Channel Coordination: Route messages between different platforms—Slack to Discord, Telegram to email—with Claude processing the content appropriately.
Security Model
ClaudeClaw's sandbox runtime inverts Docker's security approach. Instead of protecting credentials with proxies while allowing network access, it isolates processes at the OS level while keeping credentials in the host environment. This reduces attack surface while maintaining performance.
The entire system is designed to be auditable by Claude itself—every component fits within the context window, making security reviews and modifications straightforward using Claude Code's existing tooling.
