Crush

Crush is a multi-model agent (30+ models) with native MCP support and a PrepareStep pattern that lets you pre-load knowledge before the main task runs.

Setup

kcp-harness integrate crush

This generates:

crush.json

{
  "mcpServers": {
    "kcp-harness": {
      "command": "npx",
      "args": ["kcp-harness", "serve"]
    }
  },
  "prepareSteps": [
    {
      "name": "load-knowledge",
      "description": "Pre-load governed knowledge via KCP harness",
      "tool": "kcp_plan",
      "arguments": {
        "task": "{{task}}",
        "manifest": "./knowledge.yaml"
      }
    }
  ]
}

PrepareStep Pattern

The prepareSteps configuration runs kcp_plan before each task. This ensures:

1. The harness has an approved plan before Crush accesses any files
2. Plan-first mode kicks in for subsequent reads (fast path)
3. Budget is tracked from the first tool call

The placeholder is replaced with Crush's current task description.

Why kcp_plan (not kcp_load)?

The PrepareStep uses kcp_plan (not kcp_load) because:

• kcp_plan is lightweight — it returns the plan without loading content
• The plan is cached in the harness session
• Subsequent Read calls to governed paths use the cached plan (plan-first fast path)
• This avoids loading all knowledge upfront when only some units may be needed