ClaudeClaw: Composable Agent Orchestrator as a Claude Code Plugin
This week's special feature: ClaudeClaw — an open-source (MIT) agent orchestrator that runs as a Claude Code plugin, turning Claude into a persistent, always-on agent across Slack, WhatsApp, Telegram, Discord, and Gmail.
The pitch: You want Claude running 24/7 in your team's Slack — triaging support questions, investigating bugs, filing GitHub issues — not just when someone has a terminal open. ClaudeClaw makes that a `git clone` + `/setup` away.
What makes it interesting:
🔒 OS-level sandboxing, not Docker. Each agent spawns in <10ms using Anthropic's `sandbox-runtime` (Seatbelt on macOS, bubblewrap on Linux). Kernel-enforced filesystem and network restrictions. The security model inverts Docker's: credentials are passed directly, but the agent can only connect to Anthropic's API — nowhere to exfiltrate.
🧩 Composable extension system. Channels and agents aren't bundled — they're installable extensions with their own repos. `/install-extension slack` clones from GitHub, compiles, copies skills, restarts. No merge conflicts, no monolith. Core stays lean (~8K lines).
🧠 Structured memory. Daily logs, topic files, long-term CLAUDE.md, searchable archive. PreCompact hooks auto-save conversations before context compaction. QMD semantic search (hybrid BM25 + vector + LLM re-ranking) on the roadmap.
📡 Three trigger modes. Channel messages (@mention), scheduled tasks (cron), and webhook triggers (HTTP POST with HMAC-SHA256). CI failures, GitHub events, and monitoring alerts can all invoke agents.
🔧 Self-improving. The entire codebase fits in Claude's context window. Want to change behavior? Tell Claude. It reads, modifies, and commits code changes. No config files — code is the configuration.
Architecture at a glance:
```
Channels → SQLite → Polling loop → Sandbox (Claude Agent SDK) → Response
Webhooks → HTTP → HMAC verify → Queue
```
Single Node.js process. Channels self-register at startup. Per-group isolation: each group gets its own sandbox, memory, filesystem, and agent config (model, tools, cost limits).
Origin: Fork of NanoClaw, rebuilt as a Claude Code plugin in two days. Added sandbox runtime, extension system, message router with hook pipeline, structured memory, webhook triggers, cost tracking, and per-group agent configuration.
Roadmap:
- Multi-model routing (OpenAI, Google, Ollama — provider-prefixed model strings)
- QMD semantic memory (fully local, no API calls for search)
- Obsidian vault integration
- TUI for instance management
Links:
- 🔗 Repo: github.com/sbusso/claudeclaw
- 📄 License: MIT
- 📋 Requirements: macOS/Linux, Node.js 20+, Claude Code, your own API key OAuth