Before vs. after: drag to rotate, toggle to compare
Blue box = your agent. Amber boxes = infrastructure you must build and keep alive. In hosted-MCP mode they vanish: the platform runs the server, you paste one URL.
What MCP actually is
A protocol, not a product
The Model Context Protocol (open-sourced by Anthropic in Nov 2024) is a standard way for AI clients to discover and call tools, read resources, and use prompts on a server — over stdio locally or streamable HTTP remotely. Think "USB-C for AI integrations."
Hosted = remote MCP
A hosted MCP server is just a remote MCP endpoint the platform operates: it handles uptime, scaling, schema updates, and auth (typically OAuth). Your agent connects with one URL and instantly sees every tool the platform exposes — search posts, pull trends, read threads.
What dies with the bridge
The DIY path meant: write a wrapper around the REST API, deploy it somewhere, store API keys, refresh tokens, patch it when endpoints change. Four failure points, all yours. Hosted MCP moves all four to the platform's ops team.
The trade-offs
You give up control: rate limits, tool selection, and pricing are the host's call. Remote tools also widen the prompt-injection surface — a malicious post your agent reads becomes model input. Treat fetched content as untrusted data, never as instructions.
Connecting one in practice
- Find the platform's MCP endpoint, e.g.
https://mcp.example.com/mcp. - Add it to your client config — in Claude Code:
claude mcp add --transport http x https://mcp.example.com/mcp. - Authenticate via the OAuth flow the server triggers on first use.
- Your agent now lists the server's tools automatically and calls them like local functions — no SDK, no glue code.
Check yourself
Nothing. That is the point — the platform hosts the server; you add its URL to your MCP client config and authenticate.
Prompt injection. Fetched posts enter the model's context; hostile text can try to steer the agent. Mitigate with read-only scopes, output filtering, and human approval on consequential actions.
Discovery. MCP servers describe their tools (names, schemas, docs) at runtime, so the agent learns what it can do without hand-written client code for each endpoint.