Ask Claude, ChatGPT, or Cursor about your meetings — Assistly's MCP server
By The Assistly team ·
Your meeting history has always been searchable inside Assistly. The problem is that Assistly usually isn't where the work happens. You're drafting the follow-up in Claude, writing the ticket in Cursor, pulling the quarterly summary together in ChatGPT — and the thing you need is a sentence somebody said on a call three weeks ago.
So we built an MCP server. Point any AI tool at one URL, and it can read your own meetings: search them, read the recaps, pull the transcript, round up the action items. You ask in plain words, in whatever you're already using.
One meeting, end to end
Here's the shape of it, from a real afternoon — a bug bash, which is just the whole team on one call working down a list of things that need fixing.
During the call. Ninety minutes, five people. Assistly is running, filed
under that product's project. Half an hour in, someone asks why a piece of the
system was built to behave the way it does — and nobody on the call was part of
that decision. The answer is already on screen: pulled from a design session six
weeks earlier, cited back to it, while the question is still hanging in the air.
Nobody opened a doc, nobody said "let me dig that up and get back to you," and
the meeting didn't stop. That's Assistly searching the project's own history
mid-answer — the same retrieval the MCP server later exposes as ask_meetings.
The call ends. Assistly writes the recap — what was decided, and each item to fix as an action item, with an owner against it where the call made one clear.
Then the actual work. Over in Claude Code:
Pull up today's call from Assistly and start on the items assigned to me.
It calls search_meetings to find the session, get_meeting for the recap and
the action items, and — where a one-line action item isn't enough to act on —
get_meeting_transcript for what was actually said about that specific item.
Now it has the list, the reasoning behind each decision, and the constraint
somebody mentioned out loud that never made it into a ticket. It starts doing
the work.
Nothing in that sequence involved transcribing by hand, pasting a wall of transcript into a chat window, or writing the meeting up before the work could begin. The meeting was the input.
None of that is specific to engineering, either. Swap the bug list for a client call and it's the same three beats: the answer arrives while you're still on the call, the recap writes itself, and whatever you work in afterwards can read it.
Other things people ask it
Drafting a follow-up. "Write the recap email for this morning's Acme call, and include what we promised on timelines."
Picking up a thread. "When did we first talk about the migration, and what changed since?" — four calls across two months, and the words in the question don't match the words in the meeting, which is what semantic recall is for.
Answering for yourself. "What have I committed to and not done yet?"
Prepping the next one. "Pull every objection from my last five sales calls."
Why MCP, and not an integration
The Model Context Protocol is the standard AI tools now use to reach outside data. Building on it means we don't ship an Assistly plugin for Claude, then another for ChatGPT, then wait to support whatever comes next — the same server works with every client that speaks the protocol, including ones that don't exist yet.
It also means setup is genuinely one URL:
https://api.tryassistly.com/mcp
No API key, no client ID, no secret to paste into a config file and then rotate later. Your client registers itself, opens a browser, and you sign in to Assistly and click Allow. That's the whole thing. In Claude Code it's one line:
claude mcp add --transport http assistly https://api.tryassistly.com/mcp
The setup page has the exact steps for Claude, Claude Code, ChatGPT, Cursor, and Codex — pick your app and follow three steps.
What a connected app can and can't do
Worth being precise about, because you're granting an outside tool access to what was said in your meetings.
It can read. Search across transcripts, recaps, and the questions you asked your assistant mid-call. Browse your meetings by date or by project. Open a single call's recap — summary, key points, participants — or its full transcript with speakers named. Pull action items across a stretch of calls.
It can't write. There is no tool to start a session, edit a recap, delete a meeting, or change anything at all. The server is read-only by construction, not by policy — the write tools don't exist.
It only sees your account. A connection is tied to the sign-in you approved. It can't reach another user's meetings, and it can't be pointed at one by asking nicely.
You can cut it off. Every connected app is listed under Account → Connected apps with a Disconnect button. Access stops immediately — not at the end of some token's lifetime.
This is the other direction from Integrations
Assistly has had MCP support in Settings for a while, and the two are easy to mix up, so: Integrations is you giving Assistly extra tools — a web search server, your company's data source, Notion, Linear — so it can use them while answering you on a call. This is the reverse: you giving another assistant read access to your Assistly meetings.
You can run both. They don't interact, and neither one is required for the other.
Getting started
If you already use Assistly, the server is live now — pick your app on the setup page and you'll be connected before your next call.
If you don't yet: Assistly is a live meeting assistant for macOS and Windows. No bot joins your calls — it works from your computer's audio, tracks every speaker, answers questions in a private overlay while you're still on the call, and writes the recap after. The meetings it records are what the MCP server then lets every other tool you use read.