tl;dr of Fathom MCP Server

Fathom’s MCP server has nine tools, tested and verified in August 2026 against its live server.

All nine of the tools are read-only, and the MCP server functionality is available on all of its tiers, including free.

Search reads meeting titles and summaries, not transcripts. Claude couldn’t find things that were in my transcripts, but found what had made it to the summaries. Full transcripts are retrievable once you have the ID. Nothing indexes them.

Two more things. Multi-word queries need every word in the same title or summary. And it’s free to connect, but the free plan drops to basic summaries after five calls a month, which thins the only layer search can reach.

Inhaltsübersicht

Fathom’s MCP server has nine tools in total, which contradicts its own help pages, which state it has seven. Of these, all nine are read-only, and I found this after connecting Fathom’s MCP to my own Claude.

You connect the MCP server from your Fathom account to your AI assistant of choice, and your library of calls and transcripts becomes a live knowledge base you can query from a chat box.

I’ve tested Fathom a lot over the last few years, both as a general meeting recorder and for how it handles transcription in other languages, such as Swedish. It’s a solid tool, though I have found the output can be a bit patchy in places.

Which Fathom MCP Endpoint is the Official One

Before I get into how the Fathom MCP works, I just want to flag that there is another company out there called Fathom Analytics. This won’t be an issue if you’re connecting directly from the Claude or ChatGPT directories; it does mean that if you go out on the internet, or even in your own LLM, looking for information, there can be a bit of name collision.

The official Fathom MCP server is at api.fathom.ai/mcp. It’s first-party and authenticates against your existing Fathom account. Checked August 2026. There are also community-built servers wrapping Fathom’s public API. Those aren’t maintained by Fathom, they use an API key you paste in yourself, and at least one exposes a tool that creates webhooks. Read-only describes Fathom’s own server, not everything carrying the name.

The Nine Tools Fathom’s MCP Server Has

Fathom’s MCP server has nine tools total. These are two search ones, three retrieve the content, two resolve IDs, and two are just account meta details. Each of these I was able to pull live from the connected MCP in August 2026. The documentation on Fathom’s own site leaves out a couple. The likeliest explanation is that they shipped after the help page was written, rather than anything being hidden.

Fathom MCP tools, enumerated from the live server August 2026
ToolWas es tutParametersLimits
search_meetingsSearches meeting titles and summaries by keyword. Does not reach transcript text.query, recorded_by (both required), created_after, max_pagesAll words must appear in the same title or summary. Around 25 results per page, 10 pages by default, 20 maximum
find_personFinds a person by name across meeting speakers, then returns their meetings with short summaries.name, recorded_by (both required), created_afterSearches the speaker index, not transcript content
list_meetingsLists meetings with filters. Returns recording ID, title, URL, recorder and invitees.created_after, created_before, cursor, recorded_by, teams, include_summary, include_action_items, max_pagesAround 10 results per page, one page by default, 20 maximum
get_meeting_transcriptReturns the full transcript for one meeting, with speaker attribution.recording_id (required), urlSelf-limits to three transcripts per query
get_meeting_summaryReturns the AI summary for one meeting.recording_id (required)One meeting per call
get_recording_by_urlTurns a Fathom meeting URL or share link into a recording ID.url (required)Accepts full URLs and bare paths
get_recording_by_call_idTurns a numeric call ID into a recording ID.call_id (required)Call IDs and recording IDs are different numbers
list_teamsReturns the Fathom teams you belong to, for use as a filter.KeineNames only
get_identityReturns the name and email of the connected account.KeineKeine

What Fathom’s Search Actually Covers

The Fathom MCP search reads meeting titles and summaries. At this point, it does not reach the transcript. This is quicker in some respects, but in my testing I’ve run into some issues with the summaries. The most obvious one was when I did the Swedish testing; it was unable to deliver a Swedish summary in spite of transcribing the call in Swedish. Fathom transcribes in far more languages than it summarizes in, so if you work in one of the languages it doesn’t summarize, the searchable layer is in the wrong language before you start. Not a huge issue, and on the whole Fathom’s summaries are quite strong, but if you’re looking for something granular, then this search might not actually reach it.

So I ran four queries against my own library. It’s small, so I couldn’t stress test the paging limits, but the canaries turned up some interesting results.

Test 1: The Canary

First I found a word that I knew was in one of the transcripts, woodchuck, but wasn’t carried to the summary. The logic here is that the tool should be able to find that specific word on search. Fathom scanned all the meetings and wasn’t able to find the word.

fathom-mcp-search-canary-test-results

Test 2: Control

Next, I searched for a word from that same call that did make it into the summary. It was the brand name Preply, from my online French tuition. This was returned straightaway.

Same call, both words spoken out loud. One is findable, one isn’t. The only difference is whether it made the summary.

fathom-mcp-search-control-test-preply

Test 3: AND Logic

The third test was looking for the phrase review incentive program. This was not found because Fathom had split it into separate keywords. Some of the keywords appeared in the summary, but not all of them, so it returned nothing.

The error message itself shows what happened. It lists the three words separately and says no meeting matched all of them.

fathom-mcp-and-logic-keyword-split

Test 4: Person Lookup

Finally, I ran find_person on Matt, who is a labeled speaker in one of my transcripts, on a name Fathom’s own meeting list had just handed me, and on my own name as the account owner. Partial names, full names, both permission scopes. Nothing was found in any case. The empty result doesn’t report a failure either. It tells the assistant to ask when I last spoke to that person, so a lookup that found nothing arrives as a helpful follow-up question. 

 

fathom-mcp-find-person-no-speaker-found

Who Can See What: Your Meetings, Your Team’s, and Admin Access

The connector can see whatever you can see in your own Fathom account and nothing else. Fathom is very clear on this and states that the connection opens no new categories of data and can only access recordings of meetings you have permission to view. Your Fathom permission settings are ringfenced into the connection with your LLM.

What isn’t fixed is how much Claude or ChatGPT can look at. Two of the nine tools, search_meetings und find_person, require a scope on every single call and there’s no default. Your assistant picks it, either your own recordings or everything you can reach, and Fathom’s own tool descriptions tell it to go broad when the question is ambiguous. You never see the choice being made.

Ask about “what we said” versus “what I said” and you may get two different searches, with nothing telling you which ran.

On Claude team and organization plans, an admin has to switch the connector on before anyone can use it. If Fathom isn’t appearing for you, start there.

What Fathom Plan You Need for MCP

Fathom’s MCP is quite generous and is available on every plan, including the free tier. You are also able to connect MCPs using the directory on any Claude or ChatGPT tier.

fathom-mcp-claude-connector-permission

The caveats here sit around the AI. On Fathom’s free tier the first five calls each month get the full advanced summary, and that’s the part we have identified that search can reach. Anything after these gets a basic templated summary, and if the relevant information isn’t being pulled through, then you might not be able to search as efficiently as you like.

Rate Limits and How Far Back Search Reaches

These numbers come from the tool definitions themselves rather than from Fathom’s documentation. Search returns about 25 results per page and will fetch up to 20 pages. This means the hard ceiling is about 500 meetings. Unless your AI assistant decides to keep digging around in your library, it will likely reach about 250 meetings and stop searching.

Listing is smaller, at about 10 meetings per page and one page by default, up to a 20 page maximum.

Transcripts are also heavily limited, with the get_meeting_transcript tool capped at three per query, because transcripts tend to run quite long.

None of these are throttles, and none of them announce themselves. My own library is far too small to reach any of them, so this is read off the tools rather than tested. Name a date range if your library runs large.

How to Connect Fathom to Claude

I used the one in Claude’s connector directory. Go to Settings, then Connectors, search for Fathom, and approve the sign in.

You can also add api.fathom.ai/mcp as a custom connector, which needs a paid Claude plan, or add the same endpoint in Claude Code. Or skip the connector entirely and paste a transcript straight into the chat, which sidesteps every limit above.

Why Your Fathom MCP Query Returned Nothing

So you’ve connected Fathom and Claude, or ChatGPT, have asked it something and it’s come back with zilch. There are six likely reasons for this.

1. Your Search Words Are in the Transcript, Not the Summary

This is what happened when I did my testing. The search looks at the summaries only, and if the specific phrase or insight didn’t make it into that, then the recall from the AI will turn up nothing.

2. You Used More Than One Word

Every single word that you search with your query has to be in the summary. This was what happened when I was looking for the review incentive program. Some of the words were in the summary, yes, but one missing word and it gave me no response. Multi-word queries are more likely to fail than a single word.

3. Your Assistant Searched the Wrong Pool

Search needs a scope on every call, either your own recordings or everything you can access, and your assistant picks without telling you. Ask “what did we say about the renewal,” and it may search a different pool than “what did I say about the renewal.” If you mean a colleague’s call, say so in the question.”What did Dave say about the budget in June” gives the assistant a person, a topic, and a date to work with.

4. The Meeting Is Older Than the Search Reach Parameter

A search reaches about 250 meetings by default and stops without saying so. Name a date range if what you want is older.

5. The Transcript Hasn’t Finished Processing

A call that has just ended may not be in the system yet. Wait and ask again.

6. Your Assistant Is Working With a Stale Tool List

Tools are enumerated when a conversation opens, so connecting part way through a chat means the assistant may not see them. Connect first, then start a fresh chat.

How Fathom’s MCP Server Compares With Other Meeting Recorders

Fathom is by no means the only AI meeting assistant with an MCP connector, and tl;dv has its own as well. How well does Fathom’s MCP fare when compared with the capabilities of other MCP servers?

Meeting recorder MCP servers, checked against first-party documentation, August 2026
ToolRead or writeReaches beyond meetingsScoping controlsIn Claude’s directoryPlan gate
FathomRead-onlyNeinBy recorder, by teamJaEvery plan
tl;dvRead-onlyNeinBy attendee, internal or externalJaPro
FirefliesRead-onlyNeinNot documentedJaNot documented
GranolaRead-onlyNeinPersonal or workspace scopes, foldersJaEvery plan, limited on free
CirclebackRead-onlyEmail and calendarNot documentedJaNot documented
OtterRead-onlyNeinOwn and shared conversationsJaEvery plan
Kolleginnen und KollegenRead and writeKalenderPer-tool admin togglesJaNot documented

Who Fathom’s MCP Server Suits, and Who It Doesn’t

Fathom has a solid tool set, and each one is clear and has a single job. It’s available on Fathom’s free tier too, which not every tool in this category manages. It will suit you if you’re searching your own calls, your library is fairly small, and the summaries reliably capture the things you’ll go looking for later.

It’s less likely to suit you if you need to search across a team. Neither server indexes what was said on your calls, so both work by narrowing the pool and then reading transcripts, and the narrowing is where they part. Fathom scopes by who hit record. The call you want was often recorded by somebody else. tl;dv scopes by who was in the room, and filters internal against external meetings, which is the difference between narrowing to the right transcript and guessing at it.

tl;dv’s MCP server comes with Pro tier and above, is read-only, five tools, OAuth. Where it really differentiates is with team use. The ability to do attendee scoping and the internal-external filter are both built for searching other people’s calls rather than your own, which is the situation Fathom’s recorder scoping handles worst.

FAQs About Fathom MCP

Nine, enumerated from the live server on 24 August 2026. Fathom’s own help pages describe seven toggles. An April 2026 walkthrough described six and named two of them wrongly. The two missing from Fathom’s count are the ID resolvers. The server has grown steadily since launch, so every published count is a snapshot, including this one.

No. Search reads meeting titles and summaries only. Full transcripts are retrievable once you have a recording ID, so the text is reachable but nothing indexes it. I searched for a word spoken on a call and left out of its summary: no results across the whole library. A word from the same call that did reach the summary returned it immediately.

Yes, on every Fathom plan including free. The catch sits elsewhere. Fathom’s free tier caps advanced AI summaries at five calls a month, and after that you get a basic template. Since search reads summaries, the searchable layer thins out from the sixth call onward, even though recording, transcription and storage stay unlimited.

Only what you can already see in Fathom. The connector opens no new categories of data and your existing permissions carry over unchanged. Two tools require a scope on every call, either your own recordings or everything you can reach, and your assistant chooses. Nothing in the answer tells you which pool it searched, so name whose calls you mean

Most often because your words were in the transcript rather than the summary, which is all search reads. Next most likely: you used more than one word. Every word has to appear in the same title or summary, and Fathom splits your query into separate keywords requiring all of them. Try one distinctive word instead.

Search returns around 25 results per page and defaults to 10 pages, so roughly 250 meetings. It can reach 20 pages, about 500, if your assistant asks for more. Listing returns about 10 at a time. Nothing announces when a search stops short, so name a date range if your library runs large.