Typilot

Typilot's strongest asset is scope discipline: one Ctrl+Space surface that reaches every text field, plus a meeting recorder that never uploads audio. The v2.0 release in June 2026 added Meetings — mic and system-audio capture, local Whisper transcription, diarization that remembers named voices across sessions, auto-summary, and chat with the transcript. The v2.1 patch days later added in-transcript search, follow-along playback, a filler-word clean-view toggle, a custom vocabulary chip editor, and screen context via @screen with nothing extra to install. That update cadence is fast for a small local-first product, and the changelog publishes every fix rather than hiding regressions. Strengths: genuine offline operation (the only default network call the vendor names is a license check), per-action model selection across local weights (Llama, Mistral, Phi, Qwen), 27 built-in commands plus custom prompts bound to shortcuts, and image analysis through the img: command. Documentation is unusually complete for this category — fourteen chapters covering Ollama setup, model choice, permissions, history and logs, and troubleshooting when the engine won't start. The blog doubles as a buyer's hardware guide, covering RAM/VRAM math and which open-weight models fit which machine. Weaknesses and where it doesn't fit: there is no team layer. No admin dashboard, no centralized billing, no shared seat management — so it's an individual professional's tool, not a department purchase. Performance is a function of your hardware; the vendor says larger models need substantial RAM/VRAM, and machines without Apple Silicon or a capable GPU will feel it. Setup is not zero-touch: you install Ollama and choose weights, and the docs include a troubleshooting chapter precisely because local engines misbehave. Autocomplete quality also tracks model size — the changelog notes typo corrections and completions improved "especially with smaller AI models". Finally, cloud-only conveniences like web research, browsing, and shared collaboration are absent by design. Where it fits: legal, medical, and finance professionals handling privileged text; developers already running Ollama who want a system-wide front end; anyone recording sensitive calls who needs local transcription with speaker labels. Skip it if you want a hands-off assistant, need team administration, or have hardware that can't run a useful local model.