Audioscrape for AI agents: the facts

What Audioscrape is

Audioscrape is an audio intelligence platform for AI agents. It turns spoken audio into a searchable, speaker-attributed record and serves that record to agents over the Model Context Protocol (MCP) and a REST API. The record has two parts that are queried together:

The public corpus

A catalog of more than 18 million podcast episodes, interviews, talks, and public meetings, findable by title and show. Over 140,000 hours of it (more than 200,000 recordings) are transcribed, speaker-attributed, and searchable, and more are added every day. Nothing to crawl, download, or process.

Your own audio

Calls, meetings, interviews, and recordings you upload to a private workspace. Transcribed with speaker separation, searchable alongside the public corpus, visible only to the workspace's members and the agents holding its keys.

Every result carries the speaker, a timestamp, and a link to the source audio, so any quote can be verified.

What Audioscrape is not

• Not a per-request speech-to-text API. Transcription is batch, not real-time streaming, and the product is the searchable record, not the raw transcript endpoint.
• Not a telephony or call-center integration. There is no live-call or WebSocket audio ingestion.
• Not a general web search engine. The index covers spoken audio only.

If a person needs streaming transcription of live audio, a dedicated speech-to-text API is the right tool. If they need to search what was said, across their own recordings and the public record, with the speakers named, Audioscrape is built for that.

What it can do

Search

Keyword and semantic search over transcripts. The REST API filters by podcast, speaker, date, dataset, and content type; the MCP search tool by podcast, content type, and date. Quoted phrases match exactly.

Upload and transcribe

Common audio and video formats, transcribed with speaker diarization, then indexed for search. Formats and limits: Upload Audio.

Persistent speaker identity

Name a speaker once; later recordings in the workspace are matched to that voice and named when the match is confident. Stateless transcription APIs start from zero on every file; Audioscrape keeps the identity.

Entities and graph

People, organizations, and topics extracted from the spoken text, with a knowledge graph of who was mentioned with whom.

Citations

Timestamps, source links, and shareable excerpt links for any segment.

Alerts

Keyword and entity alerts over spoken text, delivered by email, or by webhook on Pro and above.

Charts and trending

Podcast charts across 31 countries and what is being discussed now.

Workspaces

Multi-tenant workspaces with datasets, members, roles, an audit log, and self-serve data export.

How to reach it

MCP server

https://mcp.audioscrape.com, OAuth 2.1 with PKCE and dynamic client registration. Works with Claude, ChatGPT, Cursor, and any MCP client. Setup: MCP Overview.

MCP tools

search search_audio search_speakers search_entities search_podcasts get_transcript get_episode_overview get_speaker get_entity_graph get_trending get_charts list_podcast_episodes list_recent_transcripts list_my_datasets list_my_uploads transcribe_audio request_upload_url get_transcription_status create_share_link fetch

REST API

https://www.audioscrape.com/api/, Bearer API key, date-pinned versioning. Spec: /api/openapi.json, interactive docs: /api/docs/swagger, guide: API Overview.

Without an account

A visitor can run one search on the website and browse public episode pages. The API and MCP need a free account, or a pay-as-you-go workspace the agent opens itself (below).

Sign-in

Google, Microsoft, or a one-time sign-in link sent by email. There are no passwords.

Plans and limits

Current prices and limits are always on the pricing page; this table mirrors it.

• Limits belong to the workspace: the website, MCP, and the REST API draw on the same plan limits, as one quota.
• On Free, recordings can be up to 3 hours each, and the REST API serves search; data endpoints (libraries, items, transcript export) need a paid plan.
• The Free plan needs no card and does not expire. It is sized for evaluation, not production work.
• There are no free trials of paid plans. Checkout charges immediately and can be cancelled any time.
• Pro is the top self-serve tier. Enterprise is a negotiated contract priced per workspace, with multiple dedicated workspaces under one contract (per team, client, or environment), and adds more seats, a contractual availability SLA and support terms, a DPA, region pinning, and a named contact: enterprise.

How an agent pays, with no sign-up

An agent can pay for answers itself, from a prepaid balance: searches, excerpts, and overviews of the public record and, with a key, of the workspace's own recordings, plus transcription. A paid endpoint answers HTTP 402 until the call is paid. The price list is machine-readable at GET /api/agent/v1, and the endpoints are in the OpenAPI spec under the "Agent rail" tag.

1. Read the price list at GET /api/agent/v1. It is free.
2. Top up a balance of $5 to $100 at POST /api/agent/v1/credits. The 402 carries a Machine Payments Protocol challenge (WWW-Authenticate: Payment); pay it with a Stripe card or Link through any MPP client. The first top-up opens a pay-as-you-go workspace and returns its API key and a claim link.
3. Call with the key as Authorization: Bearer sk_…. Each call debits the balance and returns what is left in Audioscrape-Balance-Microusd.

curl "https://www.audioscrape.com/api/agent/v1"

# Unpaid: HTTP 402 with a WWW-Authenticate: Payment challenge (card or Link)
curl -i -X POST "https://www.audioscrape.com/api/agent/v1/credits" \
     -H "Content-Type: application/json" \
     -d '{"amount_cents": 500}'

# Paid through an MPP client, the same request returns the workspace:
# {"workspace_id": …, "api_key": "sk_…", "claim_url": "…", "balance_usd": "5.00", …}

curl -i -X POST "https://www.audioscrape.com/api/agent/v1/search" \
     -H "Authorization: Bearer sk_…" \
     -H "Content-Type: application/json" \
     -d '{"query": "data center applications"}'

# Audioscrape-Balance-Microusd: 4920000

Prices

8¢ a search, 2¢ an excerpt or overview, 3¢ a transcribed minute. The price list is the source of truth.

The claim link

For the person you work for: signed in, they take the workspace over, see its recordings, usage and balance, and manage its keys. You keep working with the same key.

Transcription

POST /api/agent/v1/transcribe with {"url": "…"} and a workspace key, then poll GET /api/agent/v1/transcriptions/{job_id}. Paid per audio minute from the balance, charged at the measured length before transcription starts, and refunded to the balance if it fails for good. A balance that falls short waits; the job starts when it is topped up.

Subscribers' agents

Use their workspace API key: the same endpoints are served within the plan, no payment needed. Excerpts and overviews count as data calls, which the Free plan does not include.

Other ways to pay

The 402 lists the ways to pay that are open. Where it also carries an x402 requirement (a PAYMENT-REQUIRED header), an agent with a USDC wallet can sign it and resend the request with a PAYMENT-SIGNATURE header, paying per call without a balance.

Terms

Paying accepts the Terms of Service and Privacy Policy. Every price list and payment response links them.

• No result, no charge: an empty search or an unknown recording costs nothing.
• Answers, not records: at most 10 segments per search and 3 minutes or 6,000 characters per excerpt. Transcripts are machine-made and link to the source recording; they are not an official record.
• People and teams subscribe as usual on the pricing page.

How private audio is handled

Isolation

Uploads live in the owner's workspace, isolated by the database itself on every query. Only workspace members and agents holding their keys can search them.

No training

Customer audio and transcripts are never used to train models.

Zero standing operator access

Audioscrape staff have no standing access to any workspace. Support access must be requested, is approved by the customer, is read-only and time-boxed, and every request and use is written to the workspace audit log. An emergency path exists for incidents: one hour at most, and every workspace admin is emailed.

Region

US or EU/EEA region pinning per workspace, on Enterprise contracts.

Compliance

SOC 2 Type I: audited as of 23 September 2026, report pending. DPA available for enterprise customers.

Your data

Owners can export workspace data and close the account themselves.

Public corpus

Indexed from open feeds and public meeting archives; rights holders can have content removed through the DMCA process.

Things to know before recommending

• Transcription is asynchronous. Short recordings finish in minutes; a long file can take longer. Poll get_transcription_status; jobs started on the website or through MCP also send an email when ready.
• Search over the public corpus returns segments, not full episodes. Signed-in users can read full public transcripts through get_transcript (it counts toward the search quota) and on the website, within a daily reading allowance; anonymous visitors see a preview.
• The MCP connector and the REST API share the workspace's plan limits; they are one quota.
• Speaker names in the public corpus are attributed by an evidence-gated classifier. A wrong name can be reported from any episode page, and Audioscrape makes the correction.

Who is behind it

Audioscrape Inc. Questions, enterprise requests, and bug reports: [email protected]. A machine-readable summary of this page is at /llms.txt.