The architecture call that reaches your editor
Engineering decisions are made on calls and then live only in the code, which records the outcome and not the reasoning. Whisperer keeps the call, and over MCP your editor can read it while you are working in the file it produced.
No bot in the standup · nothing installed in your tenant
DESIGN CALL → CODE → SIX MONTHS LATER ┌───────────────┐ ┌──────────┐ ┌────────────────┐ │ three options │ │ option B │ │ why not A? │ │ two rejected │─────▶│ shipped │─────▶│ nobody recalls │ │ one constraint│ │ │ │ rewrite starts │ └───────────────┘ └──────────┘ └────────────────┘
The loss
Code preserves the decision and destroys the argument
A design review considers three approaches and picks one. The repository then contains the chosen one, cleanly, with no trace of the two that were rejected or the constraint that killed them. Six months on, a reasonable engineer looks at it, sees an odd choice, and proposes exactly the approach that was already ruled out.
An architecture decision record fixes this, in theory. In practice nobody writes one for the call that mattered.
- Call The reasoning is said out loud Options, objections and the constraint that decides it — all of it spoken, none of it written.
- Record The transcript and the map Decisions separated from discussion, each keeping the moment it came from, searchable months later.
- Editor Read from inside the work Over MCP the assistant in your editor can search those meetings without you leaving the file.
Day to day
Standups, reviews and the one call that mattered
Most engineering meetings do not need a record. The two that do are the design review and the incident post-mortem, and they are exactly the two nobody has energy to write up afterwards.
Limits
Two that engineers ask about immediately
MCP runs one way: the editor reads meetings from Whisperer. There is no repository access, no file indexing and no token for your version control.
Half the team in a conference room on one device is the case this handles worst: the room arrives through your microphone as a single undifferentiated stream.
Questions
What engineers ask first
Does anything join our calls?
No. It is a desktop application listening to your own system audio, so there is no bot in the standup and nothing for a workspace administrator to approve.
Can it write the ADR for me?
It produces the map — decisions, the reasoning around them, open questions — and that is most of the raw material. Turning it into a document your team accepts is still a human edit, and should be.
What does Cursor actually get?
Read access to the meetings you explicitly shared, within the areas you approved. It cannot delete anything and never receives your credentials.
Is there an API as well as MCP?
Yes, there is a programmatic interface documented alongside MCP. MCP is the path when you want an existing AI client to reach the data rather than writing an integration yourself.
Can I upload a finished recording?
Yes. In the Meetings section you pick the file with a button or drag it in — one at a time, up to 25 MB: mp3, m4a, wav, ogg, flac, aac, plus mp4 and webm, from which the audio is taken. From there it goes exactly like a live call: transcript, meeting map, decisions and tasks. A whole archive cannot be uploaded in one go, and in no-logs mode upload is unavailable — nothing is stored there by design.
Run it on the next design review
Then, a month later, ask why the rejected option was rejected. If the answer comes back with the moment attached, the loop is closed.