Meetily is an open-source, privacy-first AI meeting assistant from Zackriya-Solutions that runs entirely on-device, delivering live transcription, speaker diarization, and LLM-generated summaries without uploading a single second of audio to the cloud. Built in Rust on top of Tauri with a Next.js front-end, the project has reached 12,215 GitHub stars and 1,231 forks under an MIT license and ships native installers for macOS and Windows, with source builds for Linux. ## Why Local Meeting Intelligence Most meeting-assistant SaaS products (Otter, Fireflies, Read.ai) push audio to managed cloud transcribers. For regulated industries and any team that touches customer or healthcare data, that model is increasingly untenable. Meetily's README frames the pitch in concrete terms: the average data breach now costs 4.4 million dollars and GDPR enforcement has cumulatively passed 5.88 billion euros in fines. By keeping every byte of audio, every transcript, and every summary on the user's own machine, Meetily eliminates the data-exfiltration surface entirely and removes per-minute API charges from the equation. ## 4x-Faster Live Transcription with Parakeet The headline performance claim is "4x faster Parakeet/Whisper live transcription." Users can choose between OpenAI's Whisper family and NVIDIA's Parakeet TDT models depending on the accuracy-versus-latency tradeoff they want. Parakeet, with its CTC-Transducer architecture, runs especially fast on Apple Silicon thanks to CoreML and Metal acceleration, while NVIDIA CUDA, AMD ROCm, and Intel Vulkan paths cover Windows and Linux hardware. Audio capture is dual-stream: the microphone (what you say) and system audio (what you hear) are mixed intelligently so remote participants on Zoom, Meet, or Teams are transcribed as cleanly as the local speaker. ## Sortformer Speaker Diarization Meetily integrates NVIDIA's Sortformer model for real-time speaker diarization, labeling utterances by speaker as the meeting unfolds rather than as a post-processing step. Combined with timestamps, this produces transcripts that read like proper meeting minutes — "Anna (00:12:04): Let's move on to the migration plan" — and feeds the downstream summarizer enough structure to generate per-speaker action items. ## Pluggable LLM Summarization For the post-meeting summary, Meetily takes a model-agnostic stance. Out of the box it integrates with Ollama for fully-local Llama, Qwen, and Mistral inference, and it also accepts API keys for Claude, Groq, OpenRouter, and any custom OpenAI-compatible endpoint. Teams that need on-prem deployment can point Meetily at an internal vLLM or TGI server and never let a transcript leave their VPC. Summary templates cover action items, decisions, open questions, and a freeform recap. ## Tauri-Based Cross-Platform Stack The choice of Tauri over Electron keeps install size and memory footprint small, an important property for an always-on assistant. The Rust core handles the audio capture, model orchestration, and SQLite-backed history, while the Next.js front-end provides the meeting UI, timeline scrubber, and configuration panels. macOS installers are notarized and Apple Silicon-optimized; the Windows release is a standard x64 setup executable; Linux requires building from source with the provided scripts. ## Positioning and Limitations Meetily sits in the same conceptual space as proprietary tools like Granola, Krisp, and Fathom but trades managed convenience for full data sovereignty and zero per-seat fees. Limitations include the absence of a true mobile client (desktop only), the need to keep the laptop awake during meetings, and the typical accuracy gap between local Whisper-Small/Parakeet models and the largest cloud transcribers. For privacy-bound enterprises, legal teams, healthcare practitioners, and anyone tired of paying per-minute for their own conversations, those tradeoffs are increasingly acceptable.