Transcribe inside Claude and ChatGPT
The service speaks MCP over streamable HTTP at https://api.transcrib.si/mcp, with four tools — transcribe_audio, summarize_transcript, extract_actions and list_transcripts. The authorisation is OAuth 2.1 with PKCE, so a client asks the user for permission and never sees a password; the same server backs the plugins for Claude and ChatGPT.
The four tools
| Tool | Input | Output | What it is for |
|---|---|---|---|
| transcribe_audio | audio_url or audio_base64, an optional filename, language (ru or en) and diarize |
the text, the segments, the language, the provider, the duration | turn a recording into text |
| summarize_transcript | transcript_id or text, an optional language |
a summary and the key points | read a long transcript quickly |
| extract_actions | transcript_id or text |
the tasks, with the owner and the due date when they were stated | pull the tasks out of a call |
| list_transcripts | an optional limit and offset |
the list of transcripts with the id, the file name, the status and the date | find a transcript you made earlier |
How a client connects
-
Register the client
A public client registers itself with
POST /oauth/register(Dynamic Client Registration) and gets a client id; a client with a secret gets one too. -
Send the user to the authorisation page
The client opens the authorisation URL with the scope, the state and a PKCE challenge. The user signs in and approves; the browser comes back with a code.
-
Exchange the code for a token
POST /oauth/tokenwith the code, the verifier and the client id returns an access token, a refresh token and the scope. -
Call the server
POST /mcpwith the token as a bearer. The client can list the tools and call them.
The authorisation request looks like this:
GET https://api.transcrib.si/oauth/authorize
?response_type=code
&client_id=…&redirect_uri=…
&scope=transcribe+profile
&state=…&code_challenge=…&code_challenge_method=S256Claude and ChatGPT
Both plugins point at the same server. The client configuration is one JSON object:
{
"mcpServers": {
"transcrib": {
"type": "http",
"url": "https://api.transcrib.si/mcp"
}
}
}The Claude plugin is a package with a skill that explains how to use the service; the ChatGPT plugin carries the same MCP server, an onboarding skill and a review set. For the ChatGPT catalogue the domain is verified with a file at /.well-known/openai-apps-challenge: the endpoint answers with the token when it is set on the server and with 404 when it is not.
What leaves your machine
The file goes to the same API as the web app and is processed by the same engine: a Russian or a mixed file on our own infrastructure, and a file that is entirely in English possibly on a cloud provider in the United States, when that fallback is enabled on the server. The MCP server does not read your chat history, does not call other services and does not keep the conversation.
The MCP server is not a chat bot
It does not answer questions, it transcribes, summarises, pulls the tasks out of a transcript and lists what you already have. The summary it returns is a summary, not a verbatim record — when the exact wording matters, the transcript itself is one call away.
Put a file on the page and read the first lines.
The widget below is the same one as on the home page: file → transcript with timecodes and speakers → export. The transcript is saved to your account.
Drop an audio or video file
or — mp3, wav, m4a, mp4, up to 64 MB and two hours
Transcription runs under your account: sign in or create a free account.
No account yet? The free plan needs no card. Create a free account or sign in.
Questions this page answers.
Which tools does the MCP server expose?
Four: transcribe_audio, summarize_transcript, extract_actions and list_transcripts. Their names are exact, and the schemas are in the table above.
Do I need an API key for MCP?
No: the authorisation is OAuth 2.1 with PKCE, so the user's own session is what allows the call. An API key also works for a server-to-server call.
Can the assistant read my other transcripts?
list_transcripts returns the list of the transcripts of the account the token belongs to, and nothing outside it. The server does not read the chat history and does not call other services.
Does it work in ChatGPT?
The MCP server works with any MCP client today, and the plugin package for ChatGPT is written and lives in the repository. It has not been submitted to the directory yet, and the domain check at /.well-known/openai-apps-challenge still has to be answered with a token on the server before a reviewer can connect it. The page on Transcrib in ChatGPT and Claude has the current status.
What happens when the limit is reached?
The tool returns the same error codes as the REST API, including limit_exceeded. The account stays open; the call fails with a clear code.
Is a summary a verbatim record?
No. summarize_transcript returns a summary and the key points. For the exact wording, read the transcript itself — that is what transcribe_audio and list_transcripts are for.
Related pages.
The REST endpoints, the error codes and the limits.
То же самое на русском.
The general guide, with the upload widget.
Every plan, every limit, in one table.
Where the keys are created.
The live answer base: 99 questions about the price, the limits, the languages, the exports, the privacy and the API, each with the page of this site it comes from.
Mark Vesyoly, Founder.
Every number on this page comes from the service contract, the pricing page or our own measurements. The author of the page and the publisher of the site are the same person. Questions and corrections: support@transcrib.si.