MODE: HOSTED MCP
Integration model — before vs after
Hosted vs local MCP
A local MCP server is a process you run yourself (stdio transport) — great for filesystem or dev-tool access. A hosted (remote) MCP server runs on the provider's infrastructure over Streamable HTTP: you paste a URL into your client, authenticate, and its tools appear. No code, no deploys, no maintenance — the provider keeps schemas, auth, and uptime current.
Why standard schemas matter
Every MCP tool ships a JSON Schema for its inputs and outputs. That means:
- Any MCP client (Claude, Cursor, Copilot…) can call any server — N + M integrations instead of N × M.
- The model sees typed parameters, so it fills arguments correctly instead of guessing at a bespoke REST API.
- Tool discovery is dynamic: the client lists tools at connect time via
tools/list.
Real caveats
- Pricing: hosted access still meters the underlying X API — heavy reads can cost real money.
- Rate limits: an eager agent loop can burn a quota in minutes; cap tool calls.
- Permissions: OAuth scopes decide what the agent can touch. Grant read-only when you only need reads.
- Trust: tool output is untrusted content — guard against prompt injection hiding in posts.
Connect step-through
Add the server URL
Paste https://mcp.x.com into your client's MCP settings. The bridge deck locks into place.
OAuth handshake
Your client opens X's consent screen; you approve scopes. The gate on the bridge unlocks.
Tools appear
The client calls tools/list and search / thread / user tools show up next to your agents — ready to call.