Audioscrape
Audioscrape gives AI agents a searchable memory of what was said. It transcribes calls, meetings and recordings with speaker labels and timestamps, and makes them searchable alongside a public library of podcasts and public meetings. Agents get ranked, cited segments over MCP or REST.
Hosted MCP Server
npx add-mcp 'https://mcp.audioscrape.com'Installs into Claude Code, Codex, Cursor and more
Documentation
Your audio and the world's, queried together.
Most tools give you a transcript, or search public content. Audioscrape is the layer underneath both: one question, answered across what your team said and what the world said.
Upload your calls, meetings & recordings
Sales calls, support lines, interviews, field recordings. Transcribed, speaker-attributed, isolated to your workspace, searchable only by you and your agents, never used to train a model. Try an upload →
18 million episodes of public conversation
Podcasts, interviews, talks and public meetings, discoverable by title and topic — with over 100,000 hours fully transcribed, speaker-attributed and entity-linked, growing daily. Ground an answer, cite the source with a timestamp.
One search_audio_content() call spans both. Your agent doesn't care where the memory lives.
Raw audio in, agent-ready memory out.
The same pipeline runs on your uploads and the public corpus. Nothing to configure.
STEP 01
Ingest
Upload a file, point us at a URL, or search the corpus we continuously crawl. Any common audio format.
upload APIfrom-URL
STEP 02
Transcribe & attribute
State-of-the-art transcription with speaker diarization and entity linking. Segments tied to a speaker, a timestamp, and the topics they mention.
speaker attributionentitiestimestamps
STEP 03
Retrieve
Keyword + semantic search over the index, returning ranked, cited segments your agent can quote, over REST or native MCP.
RESTMCP
Retrieval shaped for how models read.
Not a transcription API you post-process, not a meeting tool locked to one workflow. Audioscrape is the retrieval layer underneath both. Your agent gets ranked, attributed segments it can quote and cite.
01 · attribution
Pin a quote to a speaker
Find a spoken claim and tie it to the named person, with a timestamp and source. Not "someone in this episode," but who.
02 · retrieval
"What did X say about Y"
Entity- and topic-aware search across the corpus and your uploads. The exact segments, not summaries.
03 · research
Multi-query sweeps
Agents fan out related queries per entity, keyword, or semantic variant, then synthesize. Built to be called many times, fast.
Agents can even discover the layer on their own: we publish an llms.txt.
Two minutes to your first result.
One endpoint returns ranked segments with speaker, timestamp, source, and score. Bearer-auth, pinned to your workspace. The MCP server wraps the same thing for tool-calling clients.
# search your uploads + the public corpus
curl "https://www.audioscrape.com/api/search" \
-H "Authorization: Bearer $AUDIOSCRAPE_KEY" \
-H "Content-Type: application/json" \
-d '{ "query": "interest rate guidance", "limit": 5 }'
# Claude, Cursor, ChatGPT: add as a connector, sign in when prompted
connector URL https://mcp.audioscrape.com # OAuth handled by your client
# headless agents & CI: same server, API-key auth
{ "mcpServers": { "audioscrape": {
"url": "https://mcp.audioscrape.com",
"headers": { "Authorization": "Bearer ${AUDIOSCRAPE_KEY}" }
}}}
# real response shape, truncated
{ "results": [ {
"speaker": "Jerome Powell", "score": 0.94,
"text": "We are not on a preset course; we will respond to the data...",
"timestamp": { "start": 724.1, "end": 731.8 },
"podcast": { "title": "FOMC Press Conference" },
"urls": { "segment": "https://www.audioscrape.com/podcast/..." }
} ] }
Your recordings are not a public feature.
Hard tenant isolation, your data never used to train models, a DPA, and a 99.5% SLA, with a path all the way to a dedicated single-tenant instance when you need it.
Isolation, escalating to dedicated
Forced row-level security by default; a single-tenant instance for buyers who require it.
Your data stays yours
Never used to train models. Forced row-level isolation, a DPA, export or delete on request. US or EU/EEA processing region; EU/EEA residency available.
Isolated environments + 99.5% SLA
Separate Dev / Test / Prod, per-env keys and quotas, transparent usage-based overage.
Moments worth a listen.
Real segments from the corpus: who said it, when they said it, one click to the audio.
[
Well, you know, obviously there's nuance. Obviously there's two sides of it. Obviously it doesn't apply to everybody, but I'm going to act like it does. And I'm going to take a very strong side towards one of them. That's the structure of a great tweet. So something that not everybody agrees with, said as fact with conviction, and then a bunch of really strongly written copy to back it up.
The Startup Ideas Podcast · How to write viral tweets with the man who beat the X algorithm →
You know, if you use the models every day, which I do, I sort of hold myself to a standard of making something either small or big with every model that comes out, you you start to appreciate the fact that these things are are not commodities, that they have comparative ad advantage at a domain level.
The a16z Show · The State of AI: Macro, Apps, and Consumer →
The planets looked completely different to the naked eye. Whereas Venus was the brightest object in the sky after the moon, Mercury was a minuscule dot almost impossible to find if you didn't know where to look. Worst of all, this astronomer had then proceeded to misinterpret this celestial omen from the gods and sent the wrong prediction to the king. The moron.
The Ancients · Origins of Astronomy →
](https://www.audioscrape.com/podcast/the-ancients/episode/origins-of-astronomy#t=154.876)[
DSA has been around for a long time in various forms, but for all intents and purposes, this group wasn't really on the radar of a lot of people in politics, and I think the general public until this past year, starting when Zoran Mamdani won in New York, and then continuing on through what's been a really pretty remarkable run of congressional Democratic primary wins.
The Daily · The Stunning Rise of the Democratic Socialists of America →
I was offered, and I don't want to go into detail, but I was offered a vice presidency a short time ago of the United States of America, and I turned it down because I knew I could not do what I wanted to do, what I'd want to do because you got constraints.
Young and Profiting (YAP) with Hala Taha: Entrepreneurship and Self-Improvement Podcast · John Paul DeJoria: The Business Mindset That Turns Setbacks Into Billion-Dollar Success | Entrepreneurship | Lost Files →
Every person, topic, and show is its own page.
Free to browse, no key. continuously ingesting, new episodes daily. Follow a person or a topic straight into the transcripts.
18.8M
episodes
450K+
shows & feeds
80K+
speakers identified
daily
fresh ingest
Trending now
What buyers and builders ask first.
Is my uploaded audio private?
Yes. Uploads live in your workspace, isolated by row-level security enforced in the database itself. Only you, your team, and agents holding your keys can search them.
Do you train models on my data?
No. Your audio and transcripts are never used to train models. They exist to answer your queries, nothing else.
What does the free tier include?
Full search across the public corpus, API and MCP access, and your first uploads. No credit card required.
How accurate are the transcripts?
State-of-the-art speech models with speaker diarization. Every segment carries a timestamp and a link to the source audio, so any quote can be verified in seconds.
Where is my data processed?
Your choice of US or EU/EEA processing region, pinned per workspace. Enterprise customers get a DPA and data-export or deletion on request.
How do agents connect?
A native MCP server plus a REST API; one key works for both. Claude, Cursor, ChatGPT, or anything that speaks MCP or HTTP.
Is Audioscrape an audio intelligence platform?
Yes. It turns raw audio into a searchable, speaker-attributed record and serves it to AI agents. Three things set it apart: a public corpus of 18M+ episodes that is already indexed, so there is no pipeline to build; speaker identity that persists, so a voice enrolled once is named in every later recording; and delivery over MCP and REST, built for agents rather than dashboards.
How is it different from a transcription API like AssemblyAI or Deepgram?
A transcription API processes one file per request and returns text; the memory of who said what is your problem. Audioscrape keeps the record: your uploads and the public corpus in one searchable index, with named speakers, timestamps, and citations, queryable by your agents for as long as you keep it. Speaker enrollment, entity graphs, and alerts on spoken text come with it. Read the agent facts page.
Give your agents something to listen to.
Search 18 million episodes in the next three seconds, then connect your own audio and make it all one layer.