Fetchproof MCP Server
# @fetchproof/mcp
stdio [Model Context Protocol](https://modelcontextprotocol.io) server for the [Fetchproof](https://fetchproof.com) public-data API. Each published operation is one tool — TikTok, Instagram, Facebook (including the Ad Library), YouTube transcripts, official race results by bib, and the rest of the catalog.
You bring your own Fetchproof API key (`FETCHPROOF_API_KEY`). No key is embedded in this package. `initialize` and `tools/list` work without a key so clients and directories can inspect the tool list.
Docs: [docs.fetchproof.com/mcp](https://docs.fetchproof.com/mcp) · Free tools: [fetchproof.com/tools](https://fetchproof.com/tools) · Sign up: [app.fetchproof.com](https://app.fetchproof.com)
## Install
Node 20 or newer. No install is required if you use `npx`:
```bash
FETCHPROOF_API_KEY=fp_your_key npx -y @fetchproof/mcp
```
Or install the CLI:
```bash
npm install -g @fetchproof/mcp
```
### Claude Desktop (`claude_desktop_config.json`)
```json
{
"mcpServers": {
"fetchproof": {
"command": "npx",
"args": ["-y", "@fetchproof/mcp"],
"env": {
"FETCHPROOF_API_KEY": "fp_your_key"
}
}
}
}
```
### Cursor (`.cursor/mcp.json` or a stdio MCP entry)
```json
{
"mcpServers": {
"fetchproof": {
"command": "npx",
"args": ["-y", "@fetchproof/mcp"],
"env": {
"FETCHPROOF_API_KEY": "fp_your_key"
}
}
}
}
```
A hosted Streamable HTTP endpoint also exists at `https://api.fetchproof.com/mcp` (send the key as `Authorization: Bearer` or `x-api-key`). This package is the local stdio server: it maps tools from the published OpenAPI operations and calls the REST API directly.
## Environment
| Variable | Required | Default | Purpose |
| --- | --- | --- | --- |
| `FETCHPROOF_API_KEY` | For `tools/call` | — | Your key from [app.fetchproof.com](https://app.fetchproof.com) |
| `FETCHPROOF_BASE_URL` | No | `https://api.fetchproof.com` | API host |
| `FETCHPROOF_TIMEOUT_MS` | No | `60000` | Per-request timeout |
## Billing
Prepaid, one-time credit packs. 1 credit per call by default. Credits never expire. There is no subscription.
- Freelance: $27 for 25,000 credits
- Business: $297 for 500,000 credits
Errors, blocked or timed-out requests, invalid requests and cache hits cost nothing. A not-found or empty result is charged only where that operation's published cost says so.
Each tool description includes that operation's published credit cost, including any empty-result or outcome rules.
## What this server can do
One tool per published operation. Priority surfaces include:
- TikTok profiles, profile videos, individual videos, comments, transcripts, search
- Instagram profiles, grid posts, reels, individual posts
- Facebook pages, posts, and the Meta Ad Library
- YouTube channels, videos, comments, and transcripts
- Official race results by bib (RunSignup, Athlinks/ChronoTrack, Race Roster)
The tool list stays in sync with the published OpenAPI file in this repository. Held-back or unpublished routes are not tools.
## Development
```bash
node mcp/scripts/sync-spec.mjs # regenerate openapi/tools.json from the published spec
node mcp/test/mcp-package.test.js
npm pack --dry-run --prefix mcp
```
This package has no runtime dependencies. Logs go to stderr; stdout is JSON-RPC only (newline-delimited; Content-Length framing is also accepted).
## License
MIT
TDQS
Scored across 133 tools
Each tool is prefixed with its source platform (tiktok_, youtube_, instagram_, reddit_, etc.) and targets a distinct resource+action, so cross-platform confusion is rare. The main overlaps are localized: reddit_post_comments (GET) vs reddit_post_comments_post (POST) do the same thing and differ only by HTTP method, and youtube_search / youtube_search_hashtag / youtube_channel_search require careful reading to distinguish. Otherwise boundaries are clean.
The dominant pattern is platform_resource_action in snake_case and it holds for most of the surface. However there are visible deviations: camelCase segments embedded in names (facebook_adLibrary_ad, google_adLibrary_advertisers_search, tiktok_get_trending_feed), path-parameter-derived names (jobs_id, jobs_id_cancel_post), and HTTP-method suffixes (reddit_post_comments_post). Some tiktok-shop tools use the tiktok_ prefix rather than tiktok_shop_, blurring the namespacing.
133 tools is far beyond what an agent can reliably scan and select from in a single flat list, forcing disambiguation-by-name alone. The breadth is somewhat justified by spanning ~20 distinct platforms, but many platforms are represented by only 1-3 tools, so the count reflects aggregation rather than a tightly scoped surface.
Depth platforms (YouTube, Instagram, Facebook, TikTok) have good lifecycle coverage including search, pagination, comments and transcripts, and an account/jobs utility layer exists. But several platforms are thin dead ends: spotify has no search, soundcloud/kwai/bluesky/twitch/linktree/bilibili offer only 1-3 read tools with no search or follower lists, so agents will hit missing-operation walls on many sources.