Audience MCP (Connect AI)

Every PodHood Channel is its own remote MCP server. It exposes one read-only tool, search, over that Channel's published episodes:

https://<slug>.podhood.com/mcp

On a verified custom domain, connect to https://<custom-domain>/mcp instead. The host used for the MCP connection is also the host used in every returned citation.

Recommended path — use Connect AI

Open the Channel's public Library and click Connect AI. The dialog builds the correct server URL and searchable server name for the current host, then shows current setup instructions for 28 clients.

Do this from the public custom domain when one is Live and you want citations to keep that brand host. Opening Connect AI from the PodHood Studio produces the Channel's PodHood subdomain endpoint instead.

Choose the installation shape

The client must support a remote Streamable HTTP MCP server. Do not configure this endpoint as a local stdio command or a legacy SSE URL.

Install prompt vs. Try-it prompt

Connect AI deliberately gives you two different prompts:

• Install prompt — wires the remote MCP server into a client. It includes the generated server name, endpoint URL, transport, no-auth rule, and asks the client to confirm that the search tool loaded.
• Try-it prompt — exercises the installed tool. It asks a grounded question and requires a direct quotation with a timestamp.

Do not paste the Try-it prompt until installation has succeeded. Do not keep reusing the Install prompt as a search query.

The Try-it prompt changes with the page where Connect AI is opened:

• Library home — uses the newest available episode Q&A question, with grounded fallbacks.
• Episode page — uses that episode's Q&A question or title.
• Person/Company/Product/Topic Hub — asks what the show has said about that subject.

This makes the first call a real retrieval test rather than a generic demo.

Manual connection examples

Use the URL generated by Connect AI when possible. If a client must be configured by hand, these are the common shapes.

claude mcp add --transport http podhood-yourshow https://yourshow.podhood.com/mcp

[mcp_servers.podhood-yourshow]
url = "https://yourshow.podhood.com/mcp"

{
  "mcpServers": {
    "podhood-yourshow": {
      "url": "https://yourshow.podhood.com/mcp"
    }
  }
}

For a normal PodHood Library, the generated name is podhood-<slug>. Under active white-label, Connect AI derives a neutral name from the custom hostname (for example, library-example-com) so the config does not reintroduce PodHood branding.

What the search tool actually returns

search runs the same hybrid retrieval core as Library Search: semantic matches plus lexical matches over transcripts, titles, and descriptions, fused into ranked Episodes. Each Episode includes its best matching Segments.

The MCP response has two equivalent representations:

• Model-readable Markdown — ranked Episode sections followed by quoted Segments. Every quote begins with a citation such as [12:03](https://…?t=723000).
• Structured content — the same result as data, including episodeId, Episode URL, metadata, matched legs, Segment text, speaker, start/end time, and absolute citation URL.

A title-only match can contain an Episode with no quoted Segment. Treat it as a discovery result, not evidence for a factual answer. An empty result explicitly says that no published episode matched; the agent should not fill the gap from memory.

Tool inputs

Retrieval recipe for reliable answers

Use fast: true for an exact name, product, or phrase when semantic recall is unnecessary. Use includeText: false only to discover Episode/citation handles; without text, the result is not sufficient grounding for an answer.

Verify the connection

After installing:

1. Confirm the client shows a server with one tool named search.
2. Run the supplied Try-it prompt.
3. Confirm the client actually calls search rather than answering from general model knowledge.
4. Check that at least one result contains quoted text and a timestamp link.
5. Open the citation and verify the Library lands at the supporting moment.
6. Ask an unrelated question and accept an empty result — this is a useful anti-hallucination check.

Troubleshooting

Opening /mcp in a browser shows 405. This is expected. The server is stateless Streamable HTTP: JSON-RPC uses POST, browser preflight uses OPTIONS, and standalone GET/DELETE session operations are intentionally rejected. Test with an MCP client, not a browser address-bar request.

The client offers stdio, SSE, and HTTP. Choose remote HTTP / Streamable HTTP. There is no command to execute and no local package.

The server connects but search is missing. Recheck that the URL ends in /mcp, remove authentication fields, and reconnect. The Install prompt explicitly asks the client to confirm the tool loaded.

Search returns nothing expected. MCP sees only successfully Indexed, non-archived Episodes. Confirm the Episode is Live, test the query in public Library Search, keep fast: false, and try a shorter concept or exact name.

Results identify the Episode but contain no quote. The match may come only from the title/description. Refine the query or search within that episodeId; do not cite an unsupported title-only result as transcript evidence.

Citations use the PodHood subdomain instead of the custom domain. Reconnect using https://<custom-domain>/mcp, ideally by opening Connect AI on the custom-domain Library. Citation URLs follow the host used by the MCP request.

ChatGPT cannot add the connector. Open Plugins from the sidebar and click the + (Add) button at the top right — if it's unavailable, turn on Developer mode under Settings → Security and login (availability can depend on your plan and workspace permissions). In the New Plugin form, keep Connection on Server URL, paste the /mcp URL, and set Authentication to No authentication. Use the exact Connect AI recipe because this UI can differ from other clients.

Protocol and discovery

• Server shape: one fresh, stateless MCP server per request; no session id and no server-side conversation state.
• Transport: Streamable HTTP with JSON responses.
• Methods: POST for JSON-RPC, OPTIONS for CORS, GET/DELETE return 405.
• Access: anonymous, public, published-only, read-only.
• CORS: open for browser-based MCP clients; Authorization is accepted but ignored in v1.
• Discovery: /.well-known/mcp and /.well-known/mcp/server-card.json advertise the per-Channel pattern; each Channel's llms.txt lists its concrete endpoint.
• Agent Skills: https://podhood.com/.well-known/agent-skills/index.json publishes podhood-library-search, which teaches compatible agents when and how to use MCP, REST, or Markdown for this retrieval task.

If the integration needs deterministic HTTP calls rather than an MCP tool, use the REST API; both adapters share the same Search result and citation-building seam.

Looking for MCP access to your own channels — drafts included, for your own agents and repurposing pipelines? That is the Creator MCP, a separate authenticated server (sign in from your MCP client, or use an API key); everything on this page is the anonymous audience surface.