How to transcribe audio with Claude, ChatGPT or Cursor
Your assistant can read text, but it cannot listen. Ears fixes that: you point the assistant at a recording and it gets the transcript back, the detected language and, if you ask, subtitles. The whole setup takes about two minutes.
1. Connect Ears once
Ears is a remote MCP server, so there is nothing to install. Add this address to your assistant:
https://useears.com/mcp- Claude (claude.ai, Desktop, Cowork): Settings, Connectors, Add custom connector. Name it Ears, paste the address, press Connect and choose Create a free account. Save the recovery key you are shown.
- Claude Code: run
claude mcp add --transport http ears https://useears.com/mcp, then/mcpto sign in. - ChatGPT, Cursor, VS Code, Windsurf, Codex: add a custom MCP server with the same address. The docs have the exact steps for each.
New accounts get 30 free credits, enough for 30 minutes of audio. No card is needed.
2. Ask for the transcript
You do not need special syntax. Try it with our sample recording, which is real and free to use:
Transcribe https://useears.com/sample.wav and tell me who should receive the report.The assistant calls the transcribe tool and answers from the text: the report goes to Maria before Friday. Along with the text it gets the language, the minutes processed and the credits charged, so it can tell you what the job cost.
3. Use your own files
For a recording on the web, give the assistant a link that opens the file itself: not a page that plays it, and not something behind a login. For a file on your computer, ask the assistant to transcribe it: in Claude Code or Cowork it asks Ears for a one-time upload command (valid for 15 minutes) and runs it for you. Files can be up to 25 MB.
4. Long recordings and cost
Processing takes about 15 seconds for every minute of audio, and assistants may stop waiting on a long call. Cut recordings longer than about 10 minutes into 5-minute parts:
ffmpeg -i in.mp4 -vn -ac 1 -ar 16000 -b:a 32k -f segment -segment_time 300 -reset_timestamps 1 part_%03d.mp3One credit pays for one minute of audio (rounded up per file) and costs €0.01, so an hour-long meeting is €0.60. Failed jobs are not charged and credits never expire.
What it cannot do
- It does not take YouTube, Instagram or TikTok page links, only direct links to files you own or have the right to process.
- It does not label who is speaking. Ask the assistant to infer speakers from context and check the result.
- It is automatic speech recognition: very good on clear speech, weaker on names, numbers and overlapping voices. Review a transcript before relying on it for anything important.
Privacy
The audio and the transcript are processed in memory while the request runs and are not stored. Ears keeps only a ledger of credits. The privacy policy has the details.