Agent Marketplace
Lets users pick skills and MCP servers once and use them in any AI client through a single parent MCP server.
Overview
Every AI client keeps its own list of skills and MCP servers. Setting them up one client at a time is tedious, the lists drift out of sync, and a setup built for one provider doesn't carry over to another. An organization also has no single place to approve, grant, or revoke what its agents can use.
Approach
The AI client is configured with exactly one MCP server: the parent. Users choose skills and MCP servers in a web catalog. When their client connects, the parent signs them in through MCP's OAuth flow and exposes only what they enabled. With 3 skills and 5 MCP servers in the catalog, a user who enables 2 skills and 1 server sees exactly those 3, and the client treats them as if they were set up locally.
Architecture
The parent is a Python server built on FastMCP. Each enabled skill is wrapped as an MCP tool, and each enabled MCP server is a remote server that the parent brings in through FastMCP's mount and proxy features, so its tools show up as the parent's own. When the client connects, MCP's OAuth flow identifies the user, and the parent applies that user's selections and credentials from the React / Next.js catalog.
Outcome
The proof of concept worked end to end. A user enabled a skill or server in the catalog, and it showed up live in Claude (Desktop and Code) and ChatGPT / Codex, with no restart, as if it were set up locally. Turning it off removed it the same way.
Stack
FastMCP handled the MCP protocol, wrapping skills as tools, and composing remote servers into the parent, which kept the parent a small Python service. MCP's built-in OAuth flow tied each connection to a user, and the catalog is a React / Next.js app.