claude-code

Claude Code

Claude Code has first-class MCP support and a powerful hooks system (PreToolUse) that enables enforcement at the tool level.

Setup

kcp-harness integrate claude-code

This generates three files:

.mcp.json — MCP server config

{
  "mcpServers": {
    "kcp-harness": {
      "command": "npx",
      "args": ["kcp-harness", "serve"]
    }
  }
}

.claude/settings.json — PreToolUse hooks

The hooks intercept Read, Edit, Write, Glob, and Grep calls. If the target path is in a governed domain, the hook blocks the call and instructs the agent to use kcp_load instead.

CLAUDE.md — Agent instructions

Instructs Claude Code to:

1. Always use kcp_load before reading governed files
2. Call kcp_plan first to see available units
3. Use harness_status to check governance state

Alternative: Manual MCP setup

claude mcp add --scope project kcp-harness -- npx kcp-harness serve

How It Works

Claude Code → Read("docs/api.md")
  │
  ├── PreToolUse hook fires
  │   └── Path "docs/" is governed → BLOCK
  │       "Use kcp_load to access governed knowledge at docs/api.md"
  │
  └── Claude Code calls kcp_load instead
      └── Harness: 14-gate cascade → approved → returns content

The hooks ensure governance even if the model "forgets" to use kcp_load. This is the SSH-guard pattern applied to knowledge access.

Available Tools

Once connected, Claude Code can use these MCP tools:

Tool	Description
kcp_plan	Get a deterministic load plan (no content)
kcp_load	Plan + load eligible unit content
kcp_trace	14-gate decision trace
harness_status	Current governance state
harness_budget	Session spend tracking
harness_temporal_check	Plan drift detection