Website · Download · Changelog
Mispher turns speech into clean text anywhere on macOS, then goes further - dictation cleanup, rewrite-in-place, translation, and an on-device agent that plans, uses tools, and answers. No cloud, no accounts. Nothing leaves your Mac.
Homebrew (recommended) - a signed, notarized build, no Xcode required:
brew install --cask dsaad68/tap/mispherDirect download - grab the latest signed Mispher.dmg from the
releases page (or the
website), open it, and drag Mispher to Applications.
On first launch Mispher walks you through downloading its on-device models. They live in your local Hugging Face cache and run entirely on your Mac.
Summon the radial dial with a hotkey and flick toward a mode - or press an arrow key. Every slice is remappable.
| Mode | What it does |
|---|---|
| Transcribe | Fast, accurate speech-to-text dropped straight into any focused field. Parakeet for English, Nemotron for ~40 languages. |
| Translate | Speak in one language, insert another - auto-translate everything you dictate, or a one-off into a target language. |
| Rewrite | Speak an instruction and Mispher rewrites the selected text in place. |
| Ask | Hold, speak a question, and an on-device agent plans, uses tools, and streams the answer back. |
Talk however suits you: hold to push-to-talk, tap to toggle, or go fully hands-free with recording that ends on a silence timeout. Three overlay styles keep it readable but unobtrusive, and voice modes and Ask can each use their own.
Ask runs an on-device DeepAgent - it plans, calls tools, and streams back an answer. Out of the box it can:
- 👁️ See your screen - take a screenshot and reason over what's on it
- 📝 Use Apple Notes - list, read, create, and update notes
- 📁 Work with files - read and edit files in a working folder
- 📋 Touch the clipboard - read and write the system clipboard
- 🔌 Connect MCP servers - bring your own Model Context Protocol tools
You decide which tools the agent may use and how each call is approved - approve, ask, or deny, per tool or per server.
Speech recognition, cleanup, translation, and the agent all run locally in Apple's MLX framework on Apple Silicon. There is no server to send audio to, no account to create, and nothing to opt out of
- because there's nowhere for your words to go.
Mispher uses LFM2.5 models (an 8B-A1B planner and a vision model) for the agent and dedicated speech models for transcription, all via MLX.
- macOS 26+ (Tahoe)
- Apple Silicon (arm64)
Open the app project in Xcode and build the Mispher scheme:
open app/Mispher/Mispher.xcodeprojXcode resolves the DeepAgents-swift Swift package on
first build (a one-time network fetch). Xcode 26+ is required - its build system emits MLX's Metal
shader library, which swift build does not co-locate.
MIT - see LICENSE. Copyright (c) Daniel Saad.
Mispher depends on the DeepAgents-swift framework (MIT), maintained separately and resolved as a Swift package.