Persistent agent orchestrator for Claude Code.
Multi-channel. Sandboxed. 8K lines you can actually read.
Three approaches to the same problem — different trade-offs.
| ClaudeClaw | NanoClaw | OpenClaw | |
|---|---|---|---|
| Codebase | ~8K lines | ~8K lines | ~400K lines |
| Security model | Kernel-enforced sandbox | Docker isolation | Opt-in sandboxing |
| Cold start | <10ms | ~2-5s | ~2-5s native |
| Docker required | No | Yes | Optional |
| Auditability | Fits in context window | Read in 8 minutes | Open source, complex |
| Configuration | Code changes (git revert) | Code changes | JSON config + CLI |
| Extensions | Pluggable system | None | Skills ecosystem |
| Memory | Structured + searchable | Basic | MEMORY.md + SOUL.md |
ClaudeClaw is built on NanoClaw, adding sandbox isolation, extensions, structured memory, and multi-channel routing.
Single process. Channels in, sandboxed agents out.
The kernel enforces isolation, not your application code.
Seatbelt (macOS) / bubblewrap (Linux). Not app-level permission checks — OS-level enforcement.
Agents can only reach api.anthropic.com. Credentials passed directly — nowhere to exfiltrate.
<10ms cold starts. No VMs, no containers, no daemon. Just an ordinary process with restrictions.
Each group's agent sees only its own files. Family chat can't read your work directory. Kernel-enforced.
~8K lines. One process. The entire codebase fits in Claude's context window for self-improvement.
Not an enterprise framework. Software that fits each user's exact needs. Fork it, modify it, own it.
No YAML sprawl. Want different behavior? Tell Claude to change the code. Every change is a commit you can revert.
Slack, WhatsApp, Telegram, Discord, Gmail. Channels self-register — the orchestrator connects whichever ones have credentials.
Daily logs, topic files, long-term CLAUDE.md, searchable archive. Auto-saved before context compaction. Upgradeable to QMD semantic search.
Channel messages (@mention), scheduled tasks (cron), webhook triggers (HMAC-SHA256). CI failures and monitoring alerts can invoke agents directly.
Per-run logging with token usage and estimated cost. Per-group budget caps. Model selection per group — use Haiku for triage, Opus for deep work.
Install only what you use. No merge conflicts. Core updates are git pull.
Socket Mode, threads, reactions
Bug triage + SWE task queue
Discord channel
Gmail channel
QMD semantic memory
Obsidian vault sync
Terminal dashboard
macOS or Linux. Node.js 20+. Claude Code.