MCP server for AI agents

Let your agent ask what investors actually said — and cite it.

Podcast Alpha MCP is a remote Model Context Protocol server. Point any MCP client at one URL and your agent gets five research tools over a knowledge graph built from finance podcasts. Every result names who said it, timestamps the exact moment, and links back to it.

Endpoint https://pa-api.knowledgefactory.io/mcp
Connect your agent → See the five tools ↓
440+episodes analyzed
5,700+attributed statements
400+tracked predictions
240themes indexed
1,500+entities resolved

Coverage runs from April 2026 to the current snapshot and grows with each ingest. Every response carries the exact coverage_window and snapshot_version that answered it, so an agent can tell the user what the corpus does and does not reach.

The surface

Five tools, priced per call.

An agent can complete the handshake and read the full tool schemas before spending anything: initialize, ping, and tools/list are free and need no credentials.

search_statements$0.001 per call

What named investors, operators, and analysts have said on record about a company, sector, or thesis. Filter by stance — bullish, bearish, neutral, mixed, informational — by speaker, or by date. Each result marks whether the claim is firsthand or relayed from someone else.

get_predictions$0.001 per call · $0.002 with scorecard

A ledger of forward-looking calls made on record, each tracked to a resolution status: hit, miss, partial, or unresolved. Ask for the speaker scorecard and your agent can weight someone’s current claim by how their past calls actually landed. Attribution credits whoever made the call, never a co-host who just agreed.

get_catalysts$0.001 per call

Upcoming events speakers flagged on air — product launches, regulatory decisions, capacity ramps, earnings setups — with the timing window they stated. Items whose window has passed without confirmation are marked lapsed, so stale catalysts are never dressed up as live ones.

get_trending_themes$0.001 per call

Which investment themes are rising, newly emerging, fading, or steady, measured as mentions per episode in the recent window against the prior one. This index is computed across the whole corpus; it is not something an agent can reconstruct from web search.

get_coverage$0.002 per call

Which episodes covered a company, person, or event, and each episode’s ranked alpha takeaways — the claims and investor implications pulled out of hours of audio, with a link to the full Podcast Alpha article. The answer to “what should I read about X” without listening to anything.

Why not just search the web

The corpus is structured where search is not.

Attribution is structural.

A statement cannot enter the index without a named speaker, a date, and a timestamp deep link. Anything that loses its citation is dropped by the serializer rather than served uncited.

Predictions are graded, not just quoted.

Calls are tracked over time and resolved. Your agent can answer “whose macro calls have actually worked” — a question the open web cannot answer without doing the bookkeeping itself.

No transcripts leave the building.

The index holds one-sentence paraphrases, titles, and metadata. Raw transcript text is never returned by any tool, so there is no bulk-text surface to scrape.

Descriptive, never advisory.

Responses report what a named person said and when. Every payload carries a fixed disclaimer. This is research material, not a recommendation engine.

Getting in

Two ways to connect — pay per call, or a beta token.

Default

Pay per call with x402

Built for agents: no account, no signup, no token to manage. A client that speaks x402 gets a quote on its first tool call and settles in USDC through Circle Gateway. Calls cost $0.001 or $0.002 each.

How payment works →

Beta access token

The alternative if you would rather not run a wallet. A bearer token in an Authorization header; tool calls are free during the beta, rate limited to 10 per hour. Issued by hand while the beta is invite-only.

Beta access →

A call that cannot succeed is never charged. Arguments are checked and entity names resolved before any payment settles, so a typo, an unknown company, or an expired cursor comes back as an ordinary error at no cost. Unknown tool names are free to probe, too.

In practice

Questions your agent can answer once it is connected.

Known limits

What this is not.

Point your agent at it.

Config for Claude Code, Claude Desktop, and any other MCP client — plus a starter prompt you can paste straight into your agent.

Connection guide → How the corpus is built →

Podcast Alpha MCP returns attributed opinions from public recordings — not investment advice. Quotes are attributed to their speakers; inclusion in the corpus is not an endorsement of any claim. See our Terms and Privacy Policy.