panini-connector-mcp

MCP server that lets AI coding agents (Claude Desktop, Cursor, Windsurf, Claude Code, etc.) install and verify the panini-connector SDK on a customer's self-hosted site — end-to-end, without human copy-pasting.

Companion to the wizard at panini.autoscreen.ai — same code generation, same verify flow, exposed as MCP tools an agent can call.

What it does

Exposes 4 tools over the Model Context Protocol:

| Tool | What it does | |---|---| | install_panini_connector | Given (language, framework, storage), returns the install command + files to write + env vars to set. Handles Next.js App/Pages, Express, FastAPI, Django, Flask. Storage: Supabase, Postgres, custom DB (MySQL/Mongo/etc.), or in-memory. | | verify_panini_connector | Fires all 4 signed HTTP ops (ping / create_post / update_post / delete_post) at the deployed URL. Returns per-op status + diagnostics. Cleans up its own smoke-test post. | | generate_shared_secret | Returns a cryptographically-random 32-byte hex secret for SEO_CONNECTOR_SECRET. | | get_envelope_schema | Returns the raw HTTP envelope contract Panini uses. For agents implementing custom handlers in languages the SDK doesn't cover (Laravel, Rails, Go, .NET, etc.). |

Install (one line per client)

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%/Claude/claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "panini-connector": {
      "command": "npx",
      "args": ["-y", "panini-connector-mcp"]
    }
  }
}

Windows note: GUI apps don't inherit shell PATH. Use cmd as the command instead:

{
  "mcpServers": {
    "panini-connector": {
      "command": "cmd",
      "args": ["/c", "npx", "-y", "panini-connector-mcp"]
    }
  }
}

Restart Claude Desktop. You should see 4 new tools available.

Cursor

Settings → Cursor Settings → MCP → Add new MCP server:

{
  "mcpServers": {
    "panini-connector": {
      "command": "npx",
      "args": ["-y", "panini-connector-mcp"]
    }
  }
}

Windsurf

Edit ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "panini-connector": {
      "command": "npx",
      "args": ["-y", "panini-connector-mcp"]
    }
  }
}

Claude Code (CLI)

claude mcp add panini-connector -- npx -y panini-connector-mcp

Example agent workflow

Once registered, prompt your AI:

"Install the panini-connector on this Next.js site. It uses Supabase — the table is blog_posts and posts render at /blog/[slug]."

The agent will:

1. Call install_panini_connector with language=node, framework=next_app, storage=supabase, existing_table=blog_posts, url_template=https://.../blog/{slug}
2. Read the returned install_command and run npm install panini-connector in your project
3. Write app/api/seo-connector/route.ts with the returned file content
4. Call generate_shared_secret and prompt you to add SEO_CONNECTOR_SECRET=<value> to your hosting env vars
5. Prompt you to redeploy
6. Once deployed, call verify_panini_connector with your URL + the secret
7. Report the result: 4 green ✓s means the round-trip works and you're ready to paste URL+secret into the Panini dashboard

Tool reference

install_panini_connector

Input:
  language: 'node' | 'python'
  framework: 'next_app' | 'next_pages' | 'express' | 'fastapi' | 'django' | 'flask'
  storage: 'supabase' | 'postgres' | 'custom_db' | 'memory'
  existing_table?: string           // default 'blog_posts'
  url_template?: string             // default 'https://yoursite.com/blog/{slug}'

Output:
  install_command: string           // shell command to run in project root
  files: [{ path, content, language }]  // files the agent writes
  env_vars: string                  // env vars to set on hosting
  post_install_note?: string        // gotchas (bodyParser: false, etc.)
  verify_command: string            // manual CLI verification command

Framework/language pairing:

• node → next_app | next_pages | express
• python → fastapi | django | flask

verify_panini_connector

Input:
  url: string                       // https://mysite.com/api/seo-connector
  secret: string                    // exact SEO_CONNECTOR_SECRET value
  timeout_ms?: number               // default 10000

Output:
  ok: boolean
  url: string
  ops: [{
    op: 'ping' | 'create_post' | 'update_post' | 'delete_post'
    status: 'pass' | 'fail'
    http_status?: number
    response_body?: unknown
    error?: string
  }]
  summary: string

generate_shared_secret

Input: {}
Output: { secret: string, length_bytes: 32 }

get_envelope_schema

Input: {}
Output: {
  http_method: 'POST'
  request_headers: { ... }
  request_body_shape: { ... }
  response_success: { ... }
  response_ping: { ... }
  response_errors: { ... }
  replay_window_seconds: 300
  hmac_pseudocode: string
  laravel_example: string
}

Use this when the customer's stack isn't Node or Python — the returned contract is enough for an agent to hand-implement the connector in any language.

Running locally / debugging

# Test the server directly (JSON-RPC over stdio)
npx panini-connector-mcp
# then type an MCP initialize request; server responds on stdout, logs on stderr

# Or with the MCP inspector
npx @modelcontextprotocol/inspector npx panini-connector-mcp

License

MIT.