OpenClaw

OpenClaw is an open-source multi-channel agent with a plugin hooks system. KCP Harness integrates via two hooks: before_prompt_build and before_agent_finalize.

Setup

kcp-harness integrate openclaw

This generates:

openclaw.json

{
  "mcpServers": {
    "kcp-harness": {
      "command": "npx",
      "args": ["kcp-harness", "serve"]
    }
  },
  "plugins": {
    "kcp-governance": {
      "hooks": {
        "before_prompt_build": {
          "tool": "kcp_load",
          "arguments": {
            "task": "{{task}}",
            "manifest": "./knowledge.yaml"
          },
          "description": "Load governed knowledge before prompt assembly"
        },
        "before_agent_finalize": {
          "tool": "kcp_trace",
          "arguments": {
            "task": "{{task}}",
            "manifest": "./knowledge.yaml"
          },
          "description": "Decision trace for audit trail before finalizing"
        }
      }
    }
  }
}

Plugin Hooks

before_prompt_build

Calls kcp_load to inject governed knowledge into the prompt before it's assembled. This ensures the model sees governed knowledge as part of its context, not as a separate tool call result.

before_agent_finalize

Calls kcp_trace after the agent produces its response, creating a 13-gate decision trace for the audit log. This provides a compliance record of what knowledge was available and why.

The Placeholder

Both hooks use , which OpenClaw replaces with the current task description. This ensures the harness plans knowledge selection based on the actual task.