Spoken

What the spokenmd/spoken MCP server does

The spokenmd/spoken MCP server connects MCP-compatible agents to Spoken, an API for retrieving transcripts from published podcasts. Retrieved transcripts are returned as Markdown with timestamps and resolved speaker names where available, making them suitable for summarization, search, RAG chunking, and other text-based workflows.

The service is focused on retrieval, not speech recognition. It does not require an agent to upload audio, run diarization, or map anonymous labels such as “Speaker 1.” Search can use ordinary text or a pasted Spotify or YouTube URL. A separate episode-listing operation can return the full back catalog for a podcast when its podcast ID is known.

Tools and capabilities

The spokenmd/spoken MCP server exposes four tools:

• search_podcasts: Finds episodes from a text query or Spotify/YouTube URL.
• list_episodes: Lists a podcast’s episodes using its podcast_id.
• get_transcript: Retrieves an episode transcript as Markdown with speaker names and timestamps.
• get_balance: Checks the remaining credit balance.

Searches and episode listings do not consume credits. The first transcript fetch costs one credit, while repeat fetches are free. Responses from the underlying API include credit usage headers, and failed calls are not charged according to the service description. The MCP tool set does not expose the API’s checkout or top-up endpoints.

Setup and configuration

Install the server through npm and configure it as a stdio MCP server:

{
  "mcpServers": {
    "spoken": {
      "command": "npx",
      "args": ["-y", "spoken-mcp"],
      "env": {
        "SPOKEN_API_KEY": "pt_your_key"
      }
    }
  }
}

The spokenmd/spoken MCP server reads credentials from SPOKEN_API_KEY. If the variable is omitted, it uses the pt_demo key. That demo key supports searching fully, but transcript access is limited to the demo episode. A real key can be obtained from spoken.md; the API uses pay-per-use credits rather than a subscription.

The repository also documents running from source with npm install, npm run build, and node dist/index.js, with the API key supplied in the environment. The package is documented for Claude Desktop, Cursor, and Cline, and the repository describes the service as usable with MCP-compatible agents.

Limitations and notes

This server depends on Spoken’s external API and an API key. It works with already-published podcasts; it is not a general-purpose speech-to-text server for uploaded audio. Transcript retrieval consumes credits on the first fetch, so agents should account for credit availability when processing many episodes. Use get_balance to inspect remaining credits before larger retrieval tasks.

The useful workflow is usually to search first, identify an episode ID, and then request its transcript. For archiving a show, an agent can list the back catalog and fetch episodes individually. Transcript quality and speaker-name resolution depend on what Spoken provides for the selected published podcast; the material does not specify coverage guarantees for every show.