Skip to content

dsaad68/mispher

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Mispher - Voice to text, and an agent. Entirely on your Mac.

Swift 6.1+ macOS 26+ Apple Silicon MIT

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.

Install

Homebrew (recommended) - a signed, notarized build, no Xcode required:

brew install --cask dsaad68/tap/mispher

Direct 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.

What it does

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.

The agent

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.

Private by design

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.

Requirements

  • macOS 26+ (Tahoe)
  • Apple Silicon (arm64)

Build from source

Open the app project in Xcode and build the Mispher scheme:

open app/Mispher/Mispher.xcodeproj

Xcode 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.

License

MIT - see LICENSE. Copyright (c) Daniel Saad.

Mispher depends on the DeepAgents-swift framework (MIT), maintained separately and resolved as a Swift package.

About

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.

Resources

License

Stars

11 stars

Watchers

0 watching

Forks

Packages

 
 
 

Contributors