A private meeting assistant with an asterisk
Meetily packages the familiar AI meeting workflow into a local desktop application. It captures microphone and system audio together, transcribes speech with Whisper or Parakeet, stores meetings in a local SQLite database, and generates a summary. It works independently of Zoom, Teams, or Meet because it records the computer's audio devices rather than joining as a bot.
The privacy case is strong when configured locally. Recordings and transcription models stay on the machine, and Ollama can produce summaries without a hosted API. That is useful for interviews, internal calls, or client conversations that should not become another cloud vendor's data.
There is an important qualifier. The README offers Claude, Groq, OpenRouter, OpenAI, and custom OpenAI-compatible providers for summaries. Choosing a hosted option sends transcript content outside the computer. Meetily can be local-first, but its privacy depends on the selected provider and endpoint. Organizations should lock that choice down instead of assuming the product name settles the data-flow question.
Easy installers, uneven platform treatment
Release 0.4.0 provides an executable and MSI for 64-bit Windows, plus a DMG and application archive for Apple Silicon macOS. For those users, installation is conventional. The next work is granting microphone and system-audio permissions, choosing devices, acquiring a transcription model, configuring a summary model or API, and running a short test. Local inference also means performance and fan noise depend on the machine.
The macOS artifact names matter: both are aarch64. Intel Mac users do not have a listed current package. Linux users get no official binary at all. The README says to clone the repository, install frontend dependencies, and run the GPU build script. Open issue #697 asks for AppImage and deb artifacts because release 0.4.0 supplies only Windows and macOS downloads.
Linux building is documented better than the short README implies. It requires Rust and Node.js plus compiler and CMake packages. The scripts detect CUDA, ROCm, Vulkan, OpenBLAS, or CPU-only operation and build a llama-helper sidecar before Tauri. Having a GPU driver is not enough: the guide says acceleration also needs the matching development toolkit. An open Arch Linux build failure and recent fixes proposed for CUDA library paths show why this path belongs to developers, not ordinary meeting participants.
The core workflow is sensibly designed
Capturing microphone and system audio is the right desktop approach. It avoids a visible meeting bot, works across conferencing services, and supports offline recordings as well as live calls. Meetily includes audio mixing with ducking and clipping prevention, which addresses the common problem of one source drowning out the other. Imported audio can be transcribed, and a beta enhancement path can rerun a recording with another model or language.
Local model choice is another strength. Whisper and Parakeet cover transcription, GPU support spans Metal, CUDA, ROCm, and Vulkan, and CPU operation remains available. For summaries, Ollama keeps the full workflow local, while a custom compatible endpoint lets an organization use infrastructure it already controls. Release 0.4 added multilingual summaries and built-in Qwen 3.5 model choices.
The product boundary is less generous than the front page initially suggests. Meetily Community Edition is the free desktop app reviewed here. The README describes Meetily Pro as a separate codebase with higher accuracy, custom templates, advanced exports, automatic meeting detection, team-oriented self-hosting, and priority support. It also says speaker identification was planned for Pro. Buyers should not read those items as features of the open repository or assume an aspiration has shipped.
Rough edges affect the actual record
Meeting software earns trust by capturing everything, not merely launching a polished window. Current reports expose risks in that chain. Issue #701 says Linux system audio can be silently absent while the recording continues with only the microphone. A proposed fix exists, but users on the released build should verify both sources in a playback. Issue #698 reports the app freezing and exiting while loading the transcription engine after an import. Another Windows report describes a clean machine failing because the Vulkan runtime was unavailable.
The summary workflow has its own constraint. Issue #677 asks for a transcription-only mode because Meetily currently always runs an LLM summary. That adds model setup and processing even for someone who only wants verbatim text. An open pull request also addresses transcript gaps across chunks in long local summaries. Until that work is released, long meetings deserve a comparison between the transcript and source audio.
Meetily does not currently offer the dependable speaker labeling expected from mature meeting-note services. The README assigns speaker identification to Pro and phrases it as future work. For interviews, panels, medical notes, or legal conversations, unlabeled or incorrectly attributed speech can be more damaging than having no automatic summary. Use manual names and review the result.
Health and the buying decision
The last push to the default branch and release 0.4.0 both landed on June 5, 2026. GitHub showed 338 open issues and pull requests, with new bug reports and contributions still arriving in August. That is substantial interest, but recent closed pull requests were generally not merged, so visible activity has not translated into a newer Community Edition release. The release history is steady from early previews through 0.4, yet the current queue and platform failures justify caution.
Documentation is above average for a young desktop AI project. The repository explains architecture, source builds, Linux GPU detection, provider choices, and Community versus Pro boundaries. Some headline wording overstates privacy by not foregrounding what remote summaries send away, and several badges and install links still refer to the repository's former meeting-minutes name.
Try Meetily if you are one person on supported hardware and local transcription is the priority. Record a disposable call first, confirm both audio channels, inspect timestamps, and compare the summary with the transcript. Teams and anyone producing an authoritative record should wait for firmer packaging, speaker handling, and reliability evidence or choose a narrower transcription tool.