Hosted MCP for social data
Add https://api.socialfetch.dev/mcp to Cursor, VS Code, or Claude, finish OAuth in the browser, and call the same 165 marketplace operations you get over REST — across 21 platforms. No API key in mcp.json.
What you get
Remote Streamable HTTP MCP server. Each operation is a typed tool (tiktok_profile_get → GET /v1/tiktok/profiles/{handle}). Arguments, credits, and envelopes match REST. Free helpers: docs_search, docs_read, auth/billing utilities.
Use MCP while building — inspect JSON, try nl_ask_post, then pin typed tools or ship with the SDK. Headless cron should use an API key over REST.
On connect, the server instructs agents: prefer typed tools over nl_ask_post, use docs_search when unsure, read lookupStatus and meta.creditsCharged — a 200 is not always found.
Connect a client
One-click install for Cursor, VS Code, and Windsurf, or paste the server URL into any MCP client that supports remote HTTP with OAuth.
Server URL: https://api.socialfetch.dev/mcp
Manual Cursor config
~/.cursor/mcp.json or project .cursor/mcp.json. Restart the client, then finish the OAuth browser flow when prompted.
More ways to connect
One-click install covers Cursor, VS Code, and Windsurf. Claude Code, Claude Desktop, and stdio-only clients need a config snippet or a CLI command instead.
VS Code — settings.json
Command Palette → MCP: Open User Configuration, or paste into workspace settings.
Claude Code — CLI
Run once, then /mcp inside Claude Code to authenticate.
Claude Desktop
Add a custom remote connector with the server URL above — Claude Desktop does not read mcp.json.
stdio-only clients
Bridge to Streamable HTTP with mcp-remote when a client only speaks stdio:
Scope the tool list
Base URL exposes every operation plus auth, billing, docs, and nl_ask_post. Large tool lists burn agent context.
Add ?platforms= or ?tools= to trim tools/list — e.g. https://api.socialfetch.dev/mcp?platforms=twitter,tiktok. tools/call still accepts any valid name. OAuth on the base URL first, then swap in a scoped URL.
Scoped connections always include auth_whoami, billing_balance_get, docs_search, docs_read, and nl_ask_post.
See it work
A typed tool call and its response envelope — trimmed here, but the shape (data, lookupStatus, meta.creditsCharged, meta.requestId) is identical to what the REST endpoint returns.
OAuth and credits
Auth is OAuth 2.1. The client opens Social Fetch, you approve access, and it stores a token. Credits charge to the signed-in account — same rates and meta.creditsCharged fields as REST. Docs tools are free.
Anonymous walk-up without OAuth can pay metered tools with x402 (USDC on Base). Whoami, balance, and nl_ask_post still need OAuth. Invalid Bearer returns 401.
Install Agent Skills
Skills teach the harness how to use Social Fetch: read llms.txt and agents.txt, connect MCP, call POST /v1/ask, and never invent routes. Pair Skills with MCP for the usual Cursor or Claude setup.
Orientation files for agents
- llms.txt
Recommended reading order, platform coverage, and links before writing integration code.
- agents.txt
Short crawl/onboarding file for agent harnesses that look for /agents.txt.
- Ask
Natural-language router (POST /v1/ask). Via MCP, call nl_ask_post, then pin the typed tool.
- MCP docs
Scoped tool URLs, client snippets, troubleshooting, and x402 notes.
- Agent Skills docs
Install flags, what the skill teaches, and Skills vs MCP.
- Cursor & Claude guide
End-to-end MCP + llms.txt workflow with copy-paste configs.
FAQ
Do I put an API key in mcp.json?
No. Use OAuth. The client opens Social Fetch in the browser, you approve, and it stores a token. Keep sfk_… keys for REST and the SDK.
How are MCP tools billed?
Endpoint tools bill like the matching REST route (usually 1 credit on a completed standard lookup). docs_search and docs_read are free. Check meta.creditsCharged on each response.
Does MCP cover all 21 platforms?
Yes. The base server exposes the marketplace catalog — 165 operations across 21 platforms — plus auth, billing, docs, and nl_ask_post helpers. Scope with ?platforms= or ?tools= when you want a smaller tools/list.
What is the difference between MCP and Agent Skills?
MCP gives live tools. Skills give durable onboarding instructions in the agent harness (read llms.txt, use Ask, never invent endpoints). Most Cursor/Claude setups use both.
How do I explore with natural language?
Call nl_ask_post (mirrors POST /v1/ask). Routing is free; the nested lookup bills normally. Read data.routedOperation, then switch to the typed MCP tool for production workflows.
Can I use MCP for headless cron?
Not well. OAuth needs a browser once per client. For unattended jobs, use an API key with REST or the TypeScript SDK. Keep MCP for interactive coding agents.
Does this work with Claude Code and Claude Desktop?
Yes. Claude Code: claude mcp add --transport http socialfetch <server URL>, then /mcp to authenticate. Claude Desktop: add a custom remote connector with the same URL — it doesn't read mcp.json.
I connected but a tool call fails — 401 or 402?
401 means OAuth didn't finish or the token expired — reconnect and approve access again. 402 means insufficient credits: top up if you're signed in, or complete the x402 payment challenge if you're calling anonymously.
Where is the install and troubleshooting detail?
The MCP docs cover one-click install, scoped URLs, Claude Code, stdio bridges, and common 401/402 fixes. This page is the product overview.
Connect MCP and start looking up
100 free credits, no card. Install the server, finish OAuth, optionally add Agent Skills so the harness stops inventing routes.