Transcribe audio to text with Claude using an MCP tool

To transcribe audio to text with Claude, connect a speech-to-text tool over MCP and give Claude a public HTTPS link to the recording. The tool turns the audio into text with word timings, and you can then ask Claude to summarize, quote, or clean up the transcript in the same conversation. With Sume's hosted MCP server at https://mcp.sume.com/mcp, Claude calls stt_create with the audio_url, waits with jobs_wait, and reads text, words, and segments from jobs_result.

Claude's side comes from Anthropic's Models overview, its Help Center article on custom connectors, and the Claude Code MCP docs; Sume's side from MCP quickstart, MCP tools and gates, and the STT 1.0 schema in the Sume API reference. All were read on 2026-09-27. Sume's basics page says hosted MCP still works but is not the primary path today; it is for clients that speak remote MCP, such as Claude, while a backend calls the speech-to-text API directly.

Why does Claude need a tool to transcribe audio?

Anthropic's models overview lists text and image input, and text output, for all current Claude models; audio is not on that list. A speech-to-text tool does the listening and hands Claude text to work with. Claude reaches such a tool over MCP, and for a custom connector Anthropic says Claude connects from its own cloud, so the tool's server must be reachable over the public internet.

How do I connect a speech-to-text tool to Claude?

Sume has no official Claude connector; you add its hosted MCP server yourself, once, then turn Write on:

• Claude, Claude Desktop, or Cowork: add https://mcp.sume.com/mcp as a custom connector (on Pro and Max, under Customize > Connectors). Add Sume to Claude as a custom connector walks through it.
• Claude Code: run claude mcp add --transport http sume https://mcp.sume.com/mcp, then claude mcp login sume to sign in.
• On Sume's consent page, switch Write on. The default sign-in is read-only (mcp:read), and paid tools, stt_create among them, return insufficient_scope until the session also has mcp:write.

What does Claude send to the speech-to-text tool?

Every STT field goes inside payload, next to a required idempotency_key. Ask Claude to show the cost first, and its first call carries dry_run: true, which returns an admission and cost preview without creating a job.

• audio_url (required): a public HTTPS link to the recording, preferably on Sume's media host.
• language_code: a hint such as en or ko. Leave it out and the language is detected.
• duration_seconds: 1–600, which sizes the usage reservation.
• segmentation: { "mode": "sentence" } adds sentence segments. Word timings need no flag; words always comes back.

{
  "idempotency_key": "episode-12-transcript",
  "dry_run": true,
  "payload": {
    "audio_url": "https://example.com/podcast/episode-12.m4a",
    "duration_seconds": 540,
    "segmentation": { "mode": "sentence" }
  }
}

How does Claude get the transcript back?

stt_create answers with a job id, not text. The transcript arrives after two more tool calls, and a long wait is split into slices rather than one long call.

Can Claude transcribe a file from my computer?

Not through this tool. Sume's docs say hosted MCP cannot read files from your laptop, and audio_url must be a public HTTPS URL. Put the recording at a public link first, such as your own site or storage, or use the Sume-hosted audio from an earlier Sume job.

What does it cost, and what are the limits?

stt_create bills $0.01 per audio minute, the STT 1.0 rate on API pricing, plus a 5.5% agent fee by default. dry_run: true shows the estimate before anything runs, and max_spend_usd caps a call when you pass it.

• Up to 10 minutes of audio per call. Split longer recordings first: Transcribe long audio files.
• No speaker labels: diarize is fixed server-side, and in current code a call that sends it is rejected.
• No live audio: the tool transcribes a finished file at a URL.