Concierge is an MCP server framework that provides progressive disclosure of tools for AI agents. It wraps an underlying MCP server and dynamically exposes only the tools relevant to the current workflow stage, helping to reduce complexity and cost while improving reliability. With Concierge, you can group tools into stages (e.g., browse, cart, checkout) and define allowed transitions between those stages to enforce your business process. Tools remain regular MCP endpoints—decorated with @app.tool() in your existing MCP server—so your prompts and tool logic stay unchanged. Concierge augments this with additional capabilities like state sharing between steps and optional semantic search to collapse large tool collections behind two meta-tools (search_tools and call_tool), making it scalable for APIs with dozens or hundreds of tools.

To use Concierge, install the concierge-sdk, wrap an existing MCP server, or scaffold a new project. You can then start the server and expose tools over multiple transports (stdio for CLI, streamable HTTP for web deployments, or SSE). The documented example shows how to declare stages and transitions, store and retrieve session state across steps, and run the server. The progressive disclosure ensures the agent never sees tools outside the current workflow context, while still allowing you to enable advanced features like semantic search when needed.

Prerequisites:\n- Python 3.9+ (recommended)\n- pip (comes with Python)\n\nInstallation steps:\n1) Create and activate a Python virtual environment (recommended):\n\n``` python -m venv venv

On macOS/Linux:

source venv/bin/activate

On Windows:

venv\Scripts\activate ````\n\n2) Install the Concierge SDK from PyPI:\n\npip install concierge-sdk\n\n3) Scaffold a new project or integrate with your existing MCP server:\n- Scaffold example:\n```

In your project directory, initialize if supported by the package tooling

(Examples in README show Concierge scaffolding; adapt to your setup)

\n4) Run your MCP server (example):\n```
python -m concierge

\nNotes:\n- If you already have an MCP server, wrap it with Concierge as shown in the README examples to gain progressive tool disclosure without changing your tool implementations.\n- Ensure your environment has network access and any needed dependencies for your tools.

Tips and considerations:\n- The environment can be extended with app.stages and app.transitions to enforce workflow logic.\n- You can enable semantic search to handle large tool sets by configuring provider_type and max_results in Config.\n- When running over HTTP, Concierge supports streamable HTTP for web deployments and a standard stdio transport for CLI-based clients.\n- If you encounter issues with tool visibility, verify that your stage names align between app.stages and app.transitions and that tools are registered with the correct decorators.\n- Use the wrap-and-go approach to start with a minimal tool set and progressively disclose more tools as needed.\n- For debugging, enable verbose logging in your Python environment and review MCP request/response traces to ensure stage/transition rules are being applied.