Can an AI agent transcribe my meeting recordings and search them later?

Yes. An AI agent connected to Audioscrape over MCP can submit a meeting recording with transcribe_audio, wait for get_transcription_status to report completion, and from then on search that meeting alongside every other recording in your workspace with search_audio, or read it in full with get_transcript. The transcript is speaker-separated and timestamped, stored in a private workspace only your account and your agents can reach, and stays searchable for as long as you keep it.

The difference between transcribing and remembering

Most transcription services process one file per request and hand back text. The next request starts from zero: the service does not know that “Speaker 2” on Tuesday’s call is the same customer as “Speaker 1” on Thursday’s, and you have to build your own store and search over the output. Audioscrape is the store. Uploads go into a workspace, get transcribed, diarized, entity-extracted and indexed, and the agent queries the record rather than re-reading files.

How to do it with Audioscrape

1. Connect the agent. Claude, ChatGPT, Cursor and any MCP client point at https://mcp.audioscrape.com; setup guides are under integrations. A headless agent uses an API key created from the workspace’s API Keys tab, sent as a Bearer token. The key is scoped to the workspace, so the agent sees the public corpus plus that workspace’s audio and nothing else.

2. Submit the recording. Ask the agent to transcribe a file, giving it an https URL the server can fetch (an object-storage presigned URL, a file on your own host). The agent calls transcribe_audio; the REST equivalent is POST /api/datasets/{id}/items with a JSON audio_url, or a multipart audio field if the file is local to the caller. Accepted formats include mp3, wav, m4a, mp4, webm, ogg and flac. A copy is stored on upload, so the source URL only needs to be reachable at that moment. You can also drag files onto your workspace dashboard by hand; web and agent uploads land in the same place.

3. Wait. Processing is batch. Short recordings finish in minutes; a long call takes longer, roughly one to three times real time depending on the queue. The agent polls get_transcription_status (REST: GET /api/items/{id}) until the item is published. An email notification is sent when it finishes.

4. Search and read. From then on:

• “What did the customer say about SSO on last week’s calls?” becomes search_audio over the workspace, returning segments with speaker and timestamp.
• “Summarize Thursday’s board call” becomes get_episode_overview (chapters, speakers, top entities) followed by get_transcript if detail is needed.
• “List what I have uploaded” is list_my_uploads; list_my_datasets shows the workspaces the key can read.

5. Name the speakers once. Open a finished transcript and name the recurring people: your team, a client, a vendor. The workspace keeps those identities, so later recordings come back with names attached automatically and the agent can filter by them.

6. Optional: alerts and webhooks. Set an alert on a phrase such as “fee waiver” or “change order” and the workspace notifies you, by email or by signed webhook on Pro and Enterprise, whenever a new recording matches.

What you get

• Speaker-separated, timestamped transcripts of your recordings, in a private workspace with row-level isolation enforced in the database.
• Full-text search across all of them, together with the public corpus if you want both in one query.
• Named speakers that persist across recordings once enrolled.
• Entities (people, organizations, products, places) extracted from each recording, with a workspace entity view.
• Team access: Pro workspaces include 5 members, Enterprise up to 10, with roles, an audit log and full export.
• Your audio and transcripts are never used to train models.

Limits and honest caveats

• Not real time. Audioscrape does not join calls or stream audio; it transcribes recordings after the fact. If you need a live meeting bot or captions, that is a different kind of tool. Record with your meeting software and upload the file, or have your agent upload it when the recording lands.
• Quota. Free accounts include 30 transcription minutes per month, a 200 MB per-file cap and 500 MB of storage, enough to try it on a couple of short calls. Regular meeting volume needs a paid plan: 300 minutes on Starter, 1,500 on Pro, custom on Enterprise. Minutes are billed on actual audio duration. There are no free trials of paid plans.
• Diarization can split or merge speakers on crosstalk and poor audio. The transcript editor lets you correct labels.
• Upload only recordings you have the right to use. Recording consent rules vary by jurisdiction and are your responsibility.
• Processing region is pinned per workspace (US or EU/EEA) for enterprise customers who need it.

Doing it without Audioscrape

Have the agent call a transcription API on each file, write the text to your own database, and build search over it. That works and gives you full control; it also means you own diarization cleanup, speaker identity across files, indexing, and access control. A workspace exists to skip that.

Plans and limits: /pricing. Facts for AI agents: /docs/for-agents.