One-Line Summary: Install Claude Code, scaffold the harness-codereview plugin directory, and write the manifest that tells Claude Code which sub-agents, hooks, slash commands, and MCP servers the plugin will provide.

Prerequisites: Step 1 (architecture overview), an Anthropic API key

Install Claude Code

If you don't already have it:

npm install -g @anthropic-ai/claude-code
claude --version

Set your API key:

export ANTHROPIC_API_KEY=sk-ant-...

Confirm Claude Code can run:

claude
> Hello!

If you see a response, you're set. Quit with /exit.

Scaffold the Plugin Directory

Pick a working directory. The plugin lives outside any specific project so you can install it into many:

mkdir -p ~/dev/harness-codereview
cd ~/dev/harness-codereview
git init

Create the plugin's directory layout:

mkdir -p .claude-plugin agents commands hooks skills mcp-server/src workers

Add a .gitignore:

cat > .gitignore <<'EOF'
node_modules/
dist/
*.log
.env
.env.local
EOF

The directory tree should now look like:

harness-codereview/
├── .claude-plugin/
├── .gitignore
├── agents/
├── commands/
├── hooks/
├── mcp-server/
│   └── src/
├── skills/
└── workers/

Write the Plugin Manifest

The manifest lives at .claude-plugin/plugin.json and tells Claude Code what the plugin provides. Create it:

{
  "name": "harness-codereview",
  "version": "0.1.0",
  "description": "A code-review swarm — three sub-agents (style, correctness, security) coordinated by hooks, plus a custom MCP tool server and a background audit worker.",
  "author": "Your Name",
  "license": "MIT",
  "provides": {
    "agents": [
      "style-reviewer",
      "correctness-reviewer",
      "security-reviewer"
    ],
    "commands": [
      "review",
      "review-strict",
      "review-explain"
    ],
    "hooks": [
      "PreToolUse",
      "PostToolUse",
      "Stop",
      "SessionStart"
    ],
    "mcpServers": [
      {
        "name": "harness-codereview-tools",
        "command": "node",
        "args": ["./mcp-server/dist/index.js"]
      }
    ]
  },
  "permissions": {
    "requires": [
      "Read",
      "Grep",
      "Bash"
    ],
    "denies": [
      "WebFetch"
    ]
  }
}

The provides section is documentation as much as wiring — it tells a user what they're installing. The permissions section declares what the plugin's hooks and sub-agents will need.

Add a Reference Settings File

Users installing the plugin need to know what settings.json configuration the plugin expects. Create settings.example.json at the plugin root:

{
  "$schema": "https://json.schemastore.org/claude-code-settings",
  "permissions": {
    "allow": ["Read", "Grep"],
    "ask": ["Edit", "Write", "Bash"],
    "deny": ["WebFetch"]
  },
  "hooks": {
    "PreToolUse": [
      { "matcher": "Bash", "command": "$CLAUDE_PLUGIN_ROOT/hooks/pre-tool-use.sh" }
    ],
    "PostToolUse": [
      { "matcher": "*", "command": "$CLAUDE_PLUGIN_ROOT/hooks/post-tool-use.sh" }
    ],
    "Stop": [
      { "command": "$CLAUDE_PLUGIN_ROOT/hooks/stop.sh" }
    ],
    "SessionStart": [
      { "command": "$CLAUDE_PLUGIN_ROOT/hooks/session-start.sh" }
    ]
  },
  "mcpServers": {
    "harness-codereview-tools": {
      "command": "node",
      "args": ["$CLAUDE_PLUGIN_ROOT/mcp-server/dist/index.js"]
    }
  }
}

$CLAUDE_PLUGIN_ROOT is a Claude Code variable that resolves to the plugin's installation directory at runtime. Using it makes the plugin portable across user installations.

Add a README Stub

Even at this early step, write a README. Future-you will thank current-you:

# harness-codereview
 
A Claude Code plugin that turns Claude Code into a code-review swarm.
 
## Install
 
```bash
/plugin marketplace add <your-git-host>/harness-codereview
/plugin install harness-codereview@harness-codereview

Use

/review                    # full review on the current change
/review-strict             # extra-strict review
/review-explain <finding>  # explain a finding in detail

What's inside

  • 3 sub-agents (style, correctness, security)
  • 4 hooks (PreToolUse, PostToolUse, Stop, SessionStart)
  • 3 slash commands
  • 1 MCP server exposing eslint / ruff / semgrep
  • 1 background audit worker (runs on session start)
 
## Verify the Skeleton
 
You should now have:

harness-codereview/ ├── .claude-plugin/plugin.json ├── .gitignore ├── README.md ├── agents/ (empty — Step 3) ├── commands/ (empty — Step 6) ├── hooks/ (empty — Step 4) ├── mcp-server/src/ (empty — Step 5) ├── settings.example.json ├── skills/ (empty) └── workers/ (empty — Step 7)

 
Commit:
 
```bash
git add .
git commit -m "Plugin skeleton: manifest, settings, README"

What This Step Did

You exercised two course concepts from Module 02:

  • settings-and-configuration-files.md — wrote settings.example.json and the plugin manifest.
  • plugin-and-marketplace-systems.md — used Claude Code's plugin layout convention.

You did not yet write any agent logic or hook code. That happens in the next steps. The skeleton is here so the rest of the work has a place to land.

Next: Step 3 - Define Three Sub-Agents →