MCP Integration

MCP Integration

An MCP server exposes tools to Claude over a standard protocol. Plugins can bundle MCPs or connect to remote ones.

Decision flow for a new domain

1. Search the MCP registry first: use search_mcp_registry with domain keywords. Examples: ["blender", "3d", "modeling"], ["figma", "design"], ["arxiv", "papers"].
2. If registry has a match: use suggest_connectors to surface it to the user for one-click connect. Wire the MCP in .mcp.json.
3. If no match: evaluate
   • (a) wrap the domain's CLI via the Bash tool (cheapest),
   • (b) write a thin MCP server that shells out to the CLI (reusable),
   • (c) write a full MCP server with a native SDK binding (highest leverage).
4. Fallback: for GUI-only tools, orchestrate via Playwright browser MCP.

.mcp.json schema

{
  "mcpServers": {
    "blender-bpy": {
      "command": "${CLAUDE_PLUGIN_ROOT}/bin/blender-mcp",
      "args": ["--blender-path", "${user_config.blender_path}"],
      "env": {
        "PYTHONPATH": "${CLAUDE_PLUGIN_DATA}/venv/lib/python3.11/site-packages"
      }
    }
  }
}

Inline vs. file

Small MCPs can go inline in plugin.json under mcpServers. Prefer .mcp.json when config gets non-trivial or you want to edit without bumping the plugin version.

Dependency install pattern

Use SessionStart hook to install deps into ${CLAUDE_PLUGIN_DATA} lazily:

{
  "hooks": {
    "SessionStart": [{
      "hooks": [{
        "type": "command",
        "command": "diff -q \"${CLAUDE_PLUGIN_ROOT}/requirements.txt\" \"${CLAUDE_PLUGIN_DATA}/requirements.txt\" >/dev/null 2>&1 || (cp \"${CLAUDE_PLUGIN_ROOT}/requirements.txt\" \"${CLAUDE_PLUGIN_DATA}/\" && python3 -m venv \"${CLAUDE_PLUGIN_DATA}/venv\" && \"${CLAUDE_PLUGIN_DATA}/venv/bin/pip\" install -r \"${CLAUDE_PLUGIN_DATA}/requirements.txt\")"
      }]
    }]
  }
}

Writing your own MCP server

• Use the official MCP SDK (TypeScript or Python) — https://modelcontextprotocol.io.
• Expose tools with tight schemas (JSONSchema for each param).
• Keep tool names verb-first and specific: blender_render_scene, not do_thing.
• Return structured JSON, not prose.
• Log to stderr; stdout is the protocol.

Anti-patterns

• Embedding API keys in mcpServers.env. Use userConfig with sensitive: true.
• Long-running tools without timeouts. Every tool call should have a sane upper bound.
• MCP servers that do 10 things. Split into focused servers per capability cluster.

References

• @../../agents-architect/references/mcp-selection.md — decision tree + registry usage

Output

When integrating MCPs for a domain:

1. Run registry search and present options.
2. If bundling: add to .mcp.json + any requirements.txt or install script.
3. If remote: document connect flow in README.
4. Add userConfig entries for any tokens/paths.
5. Validate: spin up a Claude Code session with the plugin and verify tools appear in /mcp.