Mispher
Dictate, rewrite, translate, and an agent in a single device
Start new thread

Mispher - Dictate, rewrite, translate, and an agent in a single device

by
•
On-device transcription, plus an agent that actually does things. Transcribe: Parakeet EOU (English), Parakeet TDT (25 European languages), CTC (Chinese), Nemotron (~40), Qwen3-ASR Ask: a local LFM2.5 agent that plans and calls tools, Apple Notes, clipboard, files, MCP support. Fully customizable: remap the dial, hotkeys, three HUD styles, tool approvals, swap any model or any prompt. Gemma 4, Qwen3.6 and Ornith 1.0 are coming next. No cloud, no accounts. Free, MIT. Apple Silicon, macOS 26.

Add a comment

Replies

Best
Maker
đź“Ś
Hey Product Hunt 👋 Every dictation app gives you text. Mispher gives you what you actually wanted: the text, or the paragraph rewritten, or the sentence translated, or an agent that went and did it. Agentic transcription, entirely on your Mac. No account, no telemetry. Site: Code: Hold left ⌥, flick the radial dial, and pick what happens when you let go: - Transcribe drops clean text into whatever field has focus. - Rewrite in place: highlight a sentence in any app, hold the key, say "make this shorter," and it swaps the selection where it sits. - Translate: speak one language, insert another. - Ask: a local LFM2.5 agent that plans, calls tools (Apple Notes, clipboard, files), and connects to your own MCP servers over HTTP or stdio, with OAuth when a server needs it. Writes are gated behind an approval card you set per tool. No Whisper anywhere, and no API key. Five recognizers, and you pick per language: Parakeet EOU streams English live, Nemotron streams ~40 languages, Parakeet TDT covers 25 European ones, and Parakeet CTC or Qwen3-ASR take Mandarin. Four run on the Neural Engine, Qwen3-ASR against a llama-server on localhost. The planner and vision models are MLX. Nothing about it is fixed. Remap every slice of the dial. Give each mode its own shortcut, and each shortcut its own feel: push-to-talk, tap-on-tap-off, or a long-press that goes hands-free and stops when you do. Teach it a custom dictionary so your names and jargon come out right. Decide whether releasing the key commits the text or parks it. Pick how recording appears: a Dynamic Island out of the notch, a pill under it, a draggable panel, or the full window. The agent is yours to govern too. Every tool and every MCP server gets one of three policies: approve and it runs, ask and it waits for you at an approval card, deny and the model can still see the tool but never runs it. Deactivate one entirely and it vanishes from the agent's world. Swap the model behind any mode. Turn the agent's vision on, or leave the planner blind. Coming next: Gemma 4 E4B, Qwen3.6 27B/35B, and Ornith 1.0. Ornith is the interesting one, a single 9B checkpoint that drives the planner and backs the vision subagent, so screen-aware Ask stops needing a second model download. It's free and MIT. Install with: brew install --cask dsaad68/tap/mispher Apple Silicon, macOS 26. - Daniel

On-device is the underrated headline here. In voice work the two things that actually block adoption are latency and privacy — a local model that never ships audio to a server kills both at once, and for anything with sensitive data (legal, medical, insurance) that's the difference between "can't touch it" and "ship it." Curious how accuracy holds up on proper nouns and numbers locally vs the big cloud models — that's usually where on-device stumbles. Nice to see transcription + a real agent in one place. Congrats on the launch 🚀

Maker

 Thanks a lot for your kind comment!

How does the local agent actually handle tool approvals on macOS without the cloud, does it prompt every single time or can you whitelist actions for a session?

Maker

 

Permissions are scoped per action, not just per connector. Each connector expands into its individual tools, and every tool has its own three-state permission gate: Approve (runs silently), Ask (requires human approval on every call), or Deny (blocked entirely).

So, for example, Notes is not controlled by a single switch. “Read note” and “List notes” can be set to auto-approve, while “Create note” and “Update note” remain behind an approval prompt.

How does the local agent actually decide when to call a tool versus just answering, and is that planning step visible somewhere so I can audit what it's about to do?

Maker

 

It decides based on your prompt to use which tools. Also, you have full control over which built-in tools or MCP tools are available for access.

Additionally, you can see the plan in the Ask view, similar to what is shown below.


Talking things out is so much faster than typing for me, and knowing none of it wanders off my own machine makes me way more comfortable actually leaning on it. Quietly excited about this one, Daniel.

How does the local agent actually handle tool calls that need internet access, like fetching a webpage or hitting an API, while still keeping everything on-device?

How does the on-device agent handle tool approvals when the LFM2.5 model needs to call something sensitive like Apple Notes, and can I pause approvals globally for trusted actions?

Maker

 

Permissions are scoped per action, not just per connector. Each connector expands into its individual tools, and every tool has its own three-state permission gate: Approve (runs silently), Ask (requires human approval on every call), or Deny (blocked entirely).

So, for example, Notes is not controlled by a single switch. “Read note” and “List notes” can be set to auto-approve, while “Create note” and “Update note” remain behind an approval prompt.

the rewrite-in-place feature is the one that'd actually sell me - highlight a sentence anywhere, say make this shorter, and it swaps in place. how does that hold up in apps that don't expose a normal text API, like electron apps or browser textareas buried under heavy JS overlays? does it degrade gracefully, like falling back to a clipboard paste, or does it just fail to find the selection in those cases?

Maker

 

Thanks a lot for your comment. I have tested it in the Electron apps or browser, works fine. Also, there is this auto-copy when finished behavior.

đź’Ž Pixel perfection

 good to know, and smart having the auto-copy as a safety net. one thing I'm curious about: when it falls back to clipboard instead of an in-place swap, does the app tell you that happened, or does it look identical from the user's side either way? if it's silent I could imagine someone assuming the rewrite landed in the doc when really it's just sitting on their clipboard waiting for a manual paste.

Maker

 This is a really valuable feedback. No, at the moment there is no indicator. I will add it.

 appreciate you taking it seriously. one thought on the fix itself - a subtle status-bar icon is easy to miss mid-flow since most people aren't staring at the menu bar while dictating, might be worth a brief on-screen toast right where the cursor is, same place the rewrite would've landed, so it's impossible to miss in the moment it actually matters.

The "entirely on your Mac, no telemetry" bit is exactly why I'd try this over cloud dictation — but I want to understand the agent step: when Mispher "goes and does it," is that agent running fully local against on-device models, or does the agentic action call out to an API? And is the rewrite/translate the same local model as transcription, or a separate one I'd need to pull down?

Maker

 

For Transcription (just speech-to-text), you need to select a model. There a list of models from different providers.

For rewrite and tranlate and ask mode, you need to download a Small LLM model based on your preference or based on your need, but it is suggested that you download LFM2.5-8B (the larger model, downside you need more memory) for Ask mode (agent mode) because it has better tool-calling capability.

Also, you can say for which mode and which model you want to use.

List of speech-to-text models:

List of local LLM models:


Selecting a model for rewrite mode:


Selecting a model for translate mode:


Selecting a model for Ask mode (agent mode):

  Thanks Daniel — the local-LLM download for rewrite/translate/ask makes sense. The one thing I still want to pin down for the privacy claim: the transcription models "from different providers" — do those all run on-device, or can some route audio to a cloud API? Just checking whether "no telemetry" holds across every mode, or only once you are on the downloaded local models.