OpenClaw Integration

OpenClaw is an open-source AI coding agent. WeftOS can host OpenClaw (and similar agents like NanoClaw, Codex, Cursor) as managed kernel processes with governance, sandboxing, and provenance.

How It Works

OpenClaw runs as a supervised process inside the WeftOS kernel:

WeftOS kernel
├── PID 1: OpenClaw agent (coding tasks)
├── PID 2: clawft agent (research tasks)
├── PID 3: NanoClaw agent (lightweight ops)
└── Services: governance, ExoChain, mesh, IPC

Each agent gets:

• A PID in the process table with lifecycle tracking
• Capabilities — what tools/operations it's allowed to use
• IPC — typed message passing to other agents
• Governance — every privileged action checked against the policy kernel
• ExoChain logging — cryptographic audit trail of all actions

Running OpenClaw on WeftOS

1. Start the kernel

weaver kernel start

2. Spawn OpenClaw as a managed process

weaver agent spawn --runtime openclaw --name code-reviewer

Or via the HTTP API:

POST /api/v1/execute
{
    "agent_id": "code-reviewer",
    "task": "Review the authentication module for security issues",
    "context": {
        "runtime": "openclaw",
        "workspace": "/path/to/project"
    }
}

3. Monitor

weaver kernel ps          # See running agents
weaver kernel attach      # Stream logs
weaver chain status       # Audit trail

Agent Runtime Adapters

WeftOS doesn't require agents to use the clawft pipeline. Any process that:

1. Accepts a task via stdin/IPC/HTTP
2. Executes using tools
3. Reports results

...can run as a WeftOS process. The kernel wraps it with governance and provenance regardless of the internal architecture.

Supported Runtimes

What Agents Get from WeftOS

Governance

Every tool call passes through the policy kernel. An OpenClaw agent trying to rm -rf / gets blocked by the governance engine before the command executes.

Provenance

Every action is logged to ExoChain with SHAKE-256 hash linking. You can cryptographically verify what any agent did, when, and why.

Isolation

Agents run in separate process contexts. WASM-sandboxed tools can't access the host filesystem. Capability tokens restrict what operations each agent can perform.

Memory

Agents can read/write to the shared HNSW vector store for persistent semantic memory that survives across sessions.

Mesh

In a multi-node deployment, agents on different machines communicate through the encrypted mesh. An OpenClaw agent on your laptop can delegate to a GPU-accelerated agent on a server.