Connect your own AI to your recordings: AudioMap MCP and API

The archive nobody opens again

Transcribing a meeting is easy. Making that transcript useful three weeks later is not.

What usually happens is that the material piles up: forty recordings, each with its summary, and no way to ask them anything as a whole. When the time comes to write the proposal, the minutes or the quarterly report, you open the audio again and search by hand.

AudioMap solves the first half from the start: it transcribes, separates speakers, summarises, and extracts tasks and chapters. This article is about the second half: how to connect your own AI to that material so it works with it.

There are two doors. One for your assistant, one for your programs.

The door for your assistant: MCP

MCP (Model Context Protocol) is an open standard that lets an AI assistant use external tools. If you use Claude Desktop, Cursor, LM Studio or Open WebUI with a local model, your assistant can talk to AudioMap directly.

You do not paste a transcript into the chat. You give it access, and it fetches what it needs.

In practice, you write this to your assistant:

Search my AudioMap notes for what we said about the campaign budget and draft a follow-up email for the client.

And the assistant, on its own, finds the notes where that was discussed, reads the relevant passages with their timestamps, and drafts. Without you opening AudioMap.

What your assistant can do

These are the tools AudioMap offers it:

Find and read - Search across all your notes by text, folder or tag. - List the most recent ones. - Read a whole note: summary, key points, decisions, tasks. - Read the transcript with timestamps and speakers. - See the chapters of a long recording. - Look up your highlights and markers.

Ask - Ask a specific note and get the answer with citations to the exact minute, not an unsourced paraphrase.

Bring in new material - Import a YouTube video, a podcast episode (by RSS or from Apple Podcasts) or an audio or video file from a URL. - Wait for the transcription to finish.

Generate - Generate the analysis with a specific template (minutes, sales, lecture, clinical session, journalistic interview). - Generate the chapters of a recording. - Fix a mistranscribed term across the whole transcript. - Extract an audio clip from one minute to another.

How to connect it, in three minutes

1. Sign in and create an API key in the MCP section of the dashboard.
2. Open your assistant's configuration and add the AudioMap server with that key.
3. Ask it about your notes.

The guide with the exact configuration for Claude Desktop, Cursor, LM Studio and Open WebUI is in the [MCP documentation](/docs/mcp-setup). It is five lines of configuration and nothing else to install.

The door for your programs: the API

MCP is for conversation. When what you want is automation, the REST API does the same from any language.

A real example: a firm that wants every recorded visit to end up as a draft of minutes in its document manager. Nobody is going to ask an assistant every time. You program it once:

POST /v1/notes            → imports the audio and returns the identifier
GET  /v1/notes/:id        → processing status
GET  /v1/notes/:id/transcript  → the transcript, in JSON, SRT, VTT or text
GET  /v1/notes/:id/analysis    → summary, decisions, tasks
POST /v1/notes/:id/ask         → question with citations
POST /v1/ask                   → ask ALL your notes at once
POST /v1/notes/:id/deliverables → generate the email, the minutes or the slides

It authenticates with the same key as MCP, in the `Authorization` header. The full description is published as OpenAPI, which means an LLM can read it and learn to use the API on its own.

Five use cases that already work

1. The follow-up email that writes itself. You finish a client call. You ask your assistant: "summarise today's call with Acme and write the follow-up email with the three commitments and the dates". It comes out with the names, figures and dates that were actually said, not invented.

2. Asking a whole quarter. "How many times did pricing come up in this quarter's calls, and what specific objections did they raise?" That is a question no individual summary answers, because it cuts across forty recordings.

3. A podcast turned into notes. You import the episode by its feed, and your assistant reads it and pulls out the ideas with the minute where they are said, so you only re-listen to what matters.

4. The documentation that comes out of the technical meeting. Your assistant reads the transcript of the architecture session and writes the decision with its context, as a document, ready for the repository.

5. The session with memory. A therapist, a lawyer or a doctor prepares the next session by asking the previous ones what was left open, without listening to hours of audio again.

What does not change when you connect your AI

You still decide what gets processed where. AudioMap transcribes and analyses on European servers. If you connect a local model (with LM Studio or Ollama), the content never leaves your machine for the reasoning part: AudioMap hands over the text and your model works on it locally.

Citations have a source. Every `ask` answer comes with the exact minute it came from. A claim without a citation is a claim you can reject.

Your key is yours and can be revoked. Every key is listed, named and revoked from the dashboard. No password sharing.

Where to start

If you already have an account, go to the MCP section of the dashboard, create a key and follow your assistant's guide. Ten minutes.

If you do not have one yet, you can see the product working with real recordings in the [demo](/), without signing up.

And if you would rather program against the API, the reference is on the [developers page](/developers).