v0.1.1 — in-app setup, verified downloads

First-run setup now happens inside the app, and the model download verifies itself.

Fixed

Setup no longer sends you to a terminal. The app previously failed with "python environment missing — run: meetinginsights setup", which is a dead end for anyone who does not use a shell. It now detects an unconfigured install on launch and offers a dialog explaining the one-time ~1.5 GB download, with a progress bar. Pressing Record unprepared offers the same dialog instead of an error.

A dropped connection could install a broken model. Python's read() returns empty rather than raising when a connection drops, so the download loop finished normally and a truncated 114 MB file was installed in place of the 470 MB model — and it passed the readiness check, because a truncated GGUF still begins with the right magic bytes. The download now checks its byte count against Content-Length before promoting the file, resumes with a Range request across three attempts, and the readiness check verifies size as well as magic.

Concurrent setups no longer corrupt each other. Running setup twice (the button pressed twice, or the CLI alongside the app) now fails fast on a lock instead of fighting over the same partial download.

Added

• meetinginsights ready — quiet probe, exits non-zero when setup is needed. status used to report success even on a bare install.
• meetinginsights setup --porcelain — emits STEP|/PCT|/ERR|/DONE| lines. The desktop dialog renders these rather than reimplementing setup, so the two paths cannot drift apart.

Install

sudo apt install ./meetinginsights_0.1.1_amd64.deb

Then launch MeetingInsights from the applications menu; it will offer to download the models on first run. The llama.cpp runtime is bundled, so nothing has to be compiled.

v0.1.0 — first release

MeetingInsights records a meeting, transcribes it and writes structured minutes — entirely on your machine. No account, no API key, and no audio ever leaves the device.

The Linux counterpart to the iOS app MeetingInsights: AI Minutes, sharing its icon and design language.

Install

git clone https://github.com/techbysaurabh/meetinginsights-linux
cd meetinginsights-linux && ./install.sh
meetinginsights setup      # one-time model download (~1.5 GB)

Then meetinginsights start "Sprint planning" … meetinginsights stop, or launch MeetingInsights from the applications menu.

How it works

Stage	Runs
Capture	PulseAudio — microphone to the left channel, system audio to the right
Transcribe	faster-whisper small.en, int8 on CPU, ~2.7x faster than real time
Minutes	Qwen2.5-0.5B Q4_K_M via llama.cpp

Speaker attribution comes from audio routing, not diarization. Because your microphone and everyone else's audio are separate channels, who said what is known rather than guessed — and action-item owners are read off those labels, so the language model is never asked to name a participant and cannot invent one.

Questions, decisions and commitments are classified by rules, not the model. During development a 0.5B model labelled 12 of 14 turns "action", including greetings. Replacing that step with pattern rules fixed every misclassification and cut runtime from 19.9s to 5.2s. The model only writes prose.

Requirements

Linux with PulseAudio or PipeWire, ffmpeg, Python 3.9+, GTK 3. No GPU needed. About 650 MB RAM while processing.

llama.cpp must be built once from source — upstream's prebuilt binaries need glibc 2.34, which excludes Ubuntu 20.04 and 22.04. install.sh prints the commands.

Note on recording consent

Recording a conversation without telling the other participants is illegal in many places. The tool prints a reminder when recording starts; please honour it.