Yogurt Records the Meeting Without a Bot Joining. Then It Opens One Socket.

The first thing to understand about a meeting recorder is what it puts in the call. Most of them join as a participant. A name appears in the grid, other people see it, and the recording has a visible cause. That is good for consent and bad for everything else: the bot needs admitting, it shows up in the guest list, and on some platforms it changes how the call is laid out.
Yogurt, released by Frigade under MIT, does not join. It records your microphone and whatever audio your computer is playing. No participant, no guest list change, nothing for the host to admit.
The FAQ states the consequence plainly: no bot joins the call, so let the other people on the call know you are recording. That sentence is doing more work than it looks like. When the recorder is invisible, the only consent mechanism left is you saying something.
What it runs
The pipeline is three local pieces.
Transcription uses whisper.cpp, with large-v3-turbo or a smaller model depending on the machine. On supported Macs, Apple's SpeechAnalyzer supplies live words during the call while Whisper produces the saved transcript, which is a sensible split: one engine for latency, one for the artifact you keep.
Summaries and question answering run through llama.cpp on an open-weight model from the Gemma 4, gpt-oss or Qwen3 families, picked for your hardware. The app is Electron and needs no account.
Setup downloads in two stages. A smaller notes model first, about 5 GB. Then the best one the machine can run well, which the docs put at 14 GB or more. Recording works once setup finishes; notes and questions work offline once the notes model is in. Minimum is 8 GB of memory across macOS 14.4 on Apple silicon, Windows 10 2004 or newer on x64, and Ubuntu 22.04 or similar.
Data stays in predictable places. ~/Library/Application Support/Yogurt on Mac, %APPDATA%\Yogurt on Windows, ~/.config/Yogurt or $XDG_CONFIG_HOME/Yogurt on Linux. The app goes online for model downloads and update checks, and that is the list.
The parts that are Mac only, and the parts that are not
Call detection is cross-platform but shallow. Yogurt watches which apps are using the microphone. For browser calls it also checks the tab title or a nearby calendar event with a call link. When it thinks a call started, it offers to record. Recording begins only if you accept, which is the right default for something that cannot announce itself.
Speaker names are Mac only. On Mac, with Accessibility permission, Yogurt can read who is speaking from Zoom, Teams and Meet when those apps display it. On Windows and Linux the transcript has no per-speaker attribution. If diarisation matters to you, that is the deciding line, and it is a strange place for a platform gap to sit in a 2026 product.
Calendar is also Mac only, and it reads calendars already on the computer. You do not sign into a calendar account inside Yogurt. Multi-language mode covers 99 languages for transcription, with 32 languages for the interface. Summaries follow the meeting's main language; answers follow the language of your question.
Two smaller behaviours matter. Yogurt stops recording when the call ends, by the same microphone-in-use signal. And answers across meetings come with links back to the source, which is the only feature in this category I would call non-negotiable. A summary you cannot check is a summary you have to trust.
The one socket
The interesting decision is the local MCP server.
Turn it on in Settings, AI apps, copy the helper command, and add it as a local MCP server in your assistant. Access is off by default. Connected apps get read-only access to meetings, notes and transcripts. They cannot change anything. Yogurt can also be told to stop further reads.
On Mac and Linux, the transport is a socket in your profile folder with 0600 permissions, so only your user account can open it. On Windows it is a named pipe with a new random name each launch, stored only in your profile folder.
Then comes the sentence that made me write this: there is no per-app token, so any app running under your account can read while access is on. Settings shows connected apps by their self-reported names.
Sit with that for a second. The permission model here runs per user session, not per application. Once you flip the switch, the boundary is your login, not a list of approved programs. And the app list in Settings is a convenience display, since names are self-reported.
You can read this two ways. The uncharitable reading is that a local read-only socket without per-app credentials is a permission system with the permissions filed off, and that "self-reported names" is what you write when you have not built an identity layer.
The charitable reading is that per-app tokens are hard to do well on a desktop, that a 0600 socket is a real boundary against other users on the machine, and that the alternative most people actually use is uploading transcripts to a vendor. A local socket with a documented caveat is a better deal than a cloud API with a privacy policy.
I lean toward the second reading, with one condition. The caveat has to stay in the FAQ. The moment a product like this starts describing local MCP access as secure without the qualifier, the honest engineering turns into marketing.
What it costs you
Nothing, and that is not the point. The costs are elsewhere.
Disk and memory, first. A 14 GB model plus recordings is real on a laptop that also runs a browser. The setup flow tells you if the machine is too slow, which is more graceful than most.
Attention, second. The app cannot announce itself, so consent becomes a social act you have to remember every single time. That is a worse user experience than a bot appearing in the grid, and a better one for the people on the call who did not install anything.
Platform, third. Mac gets speaker names and calendar, Windows gets a SmartScreen warning on first run because the build is not signed with a Microsoft certificate, and Linux gets the app without either of the Mac extras.
The trade the product makes is coherent. Local processing, local storage, local model, no account, and one read-only socket so your assistant can answer questions about last Tuesday. Almost every alternative gives up at least two of those.
Why it earns a place
The category has a default shape now: a bot joins, audio goes to a server, a summary comes back, and the vendor holds the transcript. Yogurt is the same product with the transport inverted, and inverting the transport forces every other decision to be made explicitly, including the ugly ones about speaker attribution and app identity.
That is the useful part. A local-first app is defined by where it writes its files and how it decides who can read them. Both are documented here in a level of detail you rarely get, down to file permissions and pipe naming, including the gap. You can disagree with the gap and still use the app, which is more than you can say for a service that never tells you what the gap is.
Installers are on the GitHub releases page for Mac arm64, Windows x64 and a Linux AppImage. The repo also documents building from source if you would rather read the code than the FAQ. Given the permission model, reading the code seems like the reasonable move.
Related articles
Google replaced the model behind Search, and the number that moved was a shell benchmark
Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.
Three ways a coding agent wastes money, measured on 30 repositories
TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.
earshot gives coding agents 922 sounds, and every one is rendered from code
An agent can write a game, an interface or a film and never hear any of it. earshot answers requests like "three variations of a footstep on gravel" from a catalogue of 922 designs built from public-domain recordings, with no audio files in the repository at all. It runs on CPU, as a CLI, a Python package, an MCP server and a Claude Code plugin.
The agents stealing API keys were not sophisticated. The gateways were open.
Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.