mcp-server

<p align="center">
  <img src="https://raw.githubusercontent.com/TranscriptFetch/mcp-server/main/assets/logo.png" alt="TranscriptFetch" width="84" height="84" />
</p>

# TranscriptFetch MCP Server

A [Model Context Protocol](https://modelcontextprotocol.io) server that gives any MCP client (Claude Desktop, Cursor, and others) access to the [TranscriptFetch API](https://transcriptfetch.com): fetch transcripts from YouTube, TikTok, Instagram and direct media file URLs, search videos, enumerate channels and playlists, and check your credit balance.

Runs locally over stdio and calls the TranscriptFetch API with your key. Prefer a hosted, remote server? Point your client at `https://transcriptfetch.com/mcp` instead (OAuth or API key).

## Tools

| Tool | What it does |
|---|---|
| `get_transcript` | Transcript for a video. YouTube, TikTok, Instagram, or a direct media URL. Set `ai_fallback: true` to transcribe the audio when no captions exist |
| `search_videos` | Search YouTube by keyword (YouTube only) |
| `list_channel_videos` | List a channel's videos (handle, ID, or URL) |
| `list_playlist_videos` | List a playlist's videos (ID or URL) |
| `get_credits` | Remaining credit balance for the key. Never billed |

Each successful fetch costs 1 credit. Failed, blocked and empty results are never charged, which matters on short-form video where many clips have no speech at all. Get a key at [the dashboard](https://transcriptfetch.com/app). Accounts start with 100 free credits and are topped back up to 100 at the start of each month.

## Install

No install needed. Run it on demand with `npx`:

```bash
TRANSCRIPTFETCH_API_KEY=tf_live_... npx -y transcriptfetch-mcp
```

Or install globally:

```bash
npm install -g transcriptfetch-mcp
```

Requires Node 18+.

### Run from source

```bash
git clone https://github.com/TranscriptFetch/mcp-server
cd mcp-server && npm install && npm run build
```

Then point your client at the built entrypoint with `"command": "node"` and
`"args": ["/absolute/path/to/mcp-server/dist/index.js"]`.

## Client configuration

### Claude Desktop

Add this to `claude_desktop_config.json` (Settings then Developer then Edit Config):

```json
{
  "mcpServers": {
    "transcriptfetch": {
      "command": "npx",
      "args": ["-y", "transcriptfetch-mcp"],
      "env": { "TRANSCRIPTFETCH_API_KEY": "tf_live_..." }
    }
  }
}
```

### Cursor

Add the same block under `mcpServers` in your Cursor MCP settings.

Restart the client, and the five tools appear.

## Example

Once connected, ask your assistant naturally:

> Get the transcript for https://youtu.be/aircAruvnKk and summarize the key points.

> Search YouTube for "how transformers work" and list the top 5 videos.

> List the latest videos from @lexfridman and pull the transcript of the newest one.

> How many TranscriptFetch credits do I have left?

The assistant picks the matching tool and works from the returned transcript or video list.

## Configuration

| Env var | Required | Default |
|---|---|---|
| `TRANSCRIPTFETCH_API_KEY` | yes | none |
| `TRANSCRIPTFETCH_BASE_URL` | no | `https://transcriptfetch.com` |

## Docker

The server speaks MCP over stdio, so there is no port to expose. `-i` is
required: without an attached stdin the transport closes immediately and the
container looks like it crashed.

```bash
docker build -t transcriptfetch-mcp .
docker run --rm -i -e TRANSCRIPTFETCH_API_KEY=tf_live_... transcriptfetch-mcp
```

## Links

- API docs: https://transcriptfetch.com/docs
- MCP docs: https://transcriptfetch.com/docs/mcp
- Node SDK: https://github.com/TranscriptFetch/node-sdk
- Python SDK: https://github.com/TranscriptFetch/python-sdk

## License

MIT