Last updated: March 2026
What is Claw-Stack?
Claw-Stack is a personal research project that wraps OpenClaw with persistent memory, multi-agent consensus, and policy enforcement β exploring how to transform a bare execution engine into a safer, more capable agent runtime.
The Problem with Raw OpenClaw
OpenClaw is a powerful execution engine for AI agents. It handles tool calls, context windows, and model routing well. But a raw OpenClaw installation gives you a stateless executor: the moment a session ends, memory is gone. There are no guardrails on what tools agents can call, no mechanism for agents to coordinate on high-stakes decisions, and no audit trail for what happened and why.
Running OpenClaw in production without additional infrastructure is like running a database without backups, transactions, or access control. It works β until something goes wrong.
What Claw-Stack Adds
Claw-Stack does not fork or modify OpenClaw. It wraps it using the Sidecar Pattern, running alongside the OpenClaw process and intercepting agent actions at defined policy gates. This means OpenClaw updates are applied cleanly without merge conflicts.
| Capability | Raw OpenClaw | Claw-Stack |
|---|---|---|
| Memory across sessions | None β context resets | 3-tier persistent memory |
| Multi-agent coordination | Manual orchestration only | Consensus protocol built-in |
| Tool access control | All or nothing | Per-agent allowlists |
| High-stakes approval gates | Not available | Human-in-the-loop workflows |
| Audit logging | None | Full action audit trail |
| Upgradeable without conflict | Yes | Yes β Sidecar pattern |
Key Value Propositions
Memory that persists
Agents remember facts, decisions, and lessons across sessions using a three-tier system: instant MEMORY.md recall, structured per-topic memory files, and a semantic vector index for deep retrieval.
Consensus before action
For high-stakes decisions β deploys, financial operations, config changes β multiple agents debate and vote before any action is executed. No single agent can unilaterally do something irreversible.
Governed by policy, not trust
Every agent operates under a declared policy: which tools it may use, which domains it may access, and which operations require human approval. Violations are blocked and logged, not ignored.
Competition-tested
The system placed #20 out of 362 teams (top 6%) at BearcatCTF 2026, solving 40 of 44 challenges autonomously in 48 hours. A real-world stress test under competition conditions.
What Claw-Stack Is Not
- βNot a fork of OpenClaw. The upstream project is untouched.
- βNot a cloud service. Claw-Stack is self-hosted and designed for air-gapped environments.
- βNot a replacement for OpenClaw. It is a runtime layer that requires OpenClaw to function.
- βNot model-specific. Claw-Stack works with any model that OpenClaw supports.
Frequently Asked Questions
What is OpenClaw?
OpenClaw is an MIT-licensed, self-hosted gateway that connects messaging apps (WhatsApp, Telegram, Discord, iMessage, and more) to AI coding agents. It runs as a Node.js daemon on your own machine and handles model calls, tool invocations, multi-agent sessions, and channel routing. Claw-Stack is a personal research project built on top of OpenClaw, adding governance and memory layers through separate modules.
Is Claw-Stack open source?
The individual modules that make up Claw-Stack are open source and self-hosted. You run everything on your own infrastructure β there is no Claw-Stack cloud and no telemetry. See each module's repository for its specific license.
Which AI models does it support?
Claw-Stack is model-agnostic at the infrastructure level. The runtime layer is tested primarily with Claude Opus 4, Claude Sonnet 4, and Claude Haiku 4 β but any model supported by OpenClaw can be used.
Authors: Qiushi Wu & Orange π