Connect an AI assistant (MCP)

show.fm runs a remote MCP server. MCP (Model Context Protocol) is the standard AI assistants use to connect to outside services, so once you connect show.fm you can ask your assistant about your own show and it reads the answer straight from your account.

Setup instructions below cover Claude (web, desktop and Claude Code), Cursor, ChatGPT and Codex.

Ask things like:

• "How did last week's episode do compared with my average?"
• "Pull the three strongest moments from the transcript of episode 42."
• "Who am I recording with tomorrow, and what did they say on the booking form?"
• "Which of my booking links has brought in the most guests overall?"

MCP access is included on every paid plan.

What it can and cannot do

The connection is read-only. It can read:

• your podcasts and the role you hold on each
• episodes, including drafts and scheduled episodes
• transcripts
• download analytics for the show and for individual episodes
• guests, upcoming recordings, and booking links

Your podcast role still applies. Owners can read download analytics. Producers can read all episodes, while cohosts can read only episodes where they have an active place on the roster, including related upcoming recordings. A transcript hidden from cohosts stays hidden, and guest biographies and private notes are available only to owners and producers.

It cannot publish or schedule an episode, edit anything, send an invitation, change a booking, or spend AI credits. There is no write access at all, so nothing your assistant does can change your account.

Two deliberate exclusions: guest email and phone fields are never sent to an AI assistant, and the connection cannot reach the guest portal.

One thing worth knowing about free text. Guest biographies, your own notes, booking form answers and transcripts are sent as written, and we cannot tell a phone number typed into a notes field from any other sentence. Obvious contact details and links carrying an access token are stripped automatically, but if you keep something private in a notes field, assume an assistant you connect can read it.

Connect your assistant

The server address is:

https://mcp.show.fm/mcp

The same address is in the app under Account Settings → AI clients, on the MCP server card, with a Copy button next to it. It never changes, and it is not a secret: an assistant still has to authorise against your account before it can read anything.

You will be asked to sign in to show.fm and approve the connection. Approve it only if you started it yourself.

Claude (claude.ai and the desktop app)

1. Open Settings → Connectors → Add custom connector.
2. Paste the server address above.
3. Follow the prompts, then approve the connection on the show.fm page that opens.

The web and desktop apps share one connection, so authorising in either covers both.

Claude Code

claude mcp add --transport http showfm https://mcp.show.fm/mcp

Then run /mcp inside Claude Code and complete the sign-in when prompted.

Cursor

Add show.fm in Settings → MCP → Add new MCP server, choosing the HTTP transport and the address above.

ChatGPT and Codex

Add show.fm as a connector in ChatGPT's developer mode, or in the Codex desktop app, using the address above. Both share one connection: authorising in either covers the other.

If connection fails after approval

If your assistant reports a connection problem after you approve, start the connection again from the assistant using https://mcp.show.fm/mcp. An expired authorisation request must also be restarted there. If you previously used the old MCP address, start a new connection setup using the address above. You do not need to disconnect another working assistant to retry.

An entry under Active connections means approval completed in show.fm; it does not confirm that the assistant finished connecting. Ask the assistant to list your podcasts to check that it can read from your account.

What you are approving

The approval screen names the assistant that asked and lists what it can read. Two things worth knowing before you approve:

• The connection acts as you, across every podcast you belong to on a paid plan. It cannot see podcasts you do not belong to.
• Your show.fm password and session are never shared with the assistant. The assistant receives its own credential, which only works for the tools listed above and which you can revoke at any time.

Disconnect an assistant

Go to Account Settings → AI clients. Under Active connections every assistant is listed with when it connected and when it last read anything, and a Disconnect button.

Disconnecting takes effect immediately, on the connection's very next request.

Disconnect all AI clients sits in the Danger zone at the foot of the page. It asks you to confirm, and names how many connections it is about to end.

Changing your password disconnects every AI client automatically, whether you change it in Account Settings or reset it with Forgot password?. If you think an assistant has been connected without your knowledge, change your password: that revokes everything in one step.

Past connections

Below the live ones, Past connections shows what has ended in the last 90 days, and why: disconnected by you, disconnected by a password change, or Unfinished.

An unfinished connection is an authorisation that was started and never completed. It never had access to anything. Seeing one you do not recognise is worth a password change, but it does not mean anything was read.

Good practice

• Connect only assistants you trust, on devices you control.
• Transcripts and guest booking answers are written by other people. A well-behaved assistant treats them as information, not as instructions, and show.fm marks them as untrusted when it sends them. Still, be wary of any assistant that suddenly proposes an action you did not ask for after reading a transcript.
• Review Account Settings → AI clients occasionally and remove anything you no longer use.

Tool reference

These are the tools your assistant can use. You never call them by name yourself: the assistant picks the right one from what you ask. They are listed so you know exactly what a connected assistant can and cannot see.

Tool	What it does
list_podcasts	Lists the podcasts you can access and the role you hold on each. Only podcasts on accounts whose plan includes MCP appear.
podcast_overview	Podcast details plus owner-only headline download figures for a period (7 days, 30 days, 90 days, 12 months, or all time). When the period starts later than asked, range.clampReason says why: retention (your plan's history window) or collection_start (no data before that day).
list_episodes	Your episodes, newest first, filterable by status: draft, scheduled, published or archived.
episode_brief	One episode in a single call: its details, the people on it, owner-only download figures, and whether a visible transcript exists.
get_transcript	Reads a transcript in segments. Long episodes come back in pieces rather than all at once.
upcoming_recordings	Confirmed and pending recordings still to come, with guest names, booking-form answers and notes.
list_guests	Guests associated with a podcast. Owners and producers also receive biographies and private notes.
list_booking_links	Your booking links with their settings and total bookings to date.

This list is a contract, not a snapshot. New tools can appear without notice, but a tool will not be renamed or removed, and its inputs will not be narrowed, without a deprecation notice first. That is the same promise the public API versioning policy makes.

Two things no tool can do: return a guest's email or phone field, and reach the guest portal. As above, free text is different: a phone number typed into a notes field is just text, and we cannot tell it apart from any other sentence.

When a tool declines

Some replies are refusals rather than data, and each carries a stable code your assistant may show you:

Code	Meaning
no_entitled_podcasts	None of your podcasts sit on an account whose plan includes MCP access.

Limits

Requests are rate limited per account. A normal conversation will never reach the limit; an assistant stuck in a loop will, and it will be told to slow down rather than being disconnected.

Transcripts are returned in segments rather than all at once, so an assistant reading a long episode fetches it in pieces.