Ghost Pepper: Local Voice Dictation for Apple Silicon Macs

Matt Hartman shipped Ghost Pepper to an Apple Silicon Mac near you, and the pitch is refreshingly short: no cloud, no account, no data leaving your machine.

Ghost Pepper is a macOS app for voice dictation and meeting transcription that runs every model locally. Hold Control, talk, release, and the transcription pastes directly into whatever text field you had focused. The meeting transcription side records calls and spits out a transcript, notes, and an AI-generated summary, all saved as local markdown files. A local LLM handles the cleanup pass, stripping filler words and fixing self-corrections before you ever see the output. The whole thing requires macOS 14.0 or later and an Apple Silicon chip (M1 or better), which is honestly a reasonable ask given that it’s doing actual inference on your device.

The GitHub repo has 2,100 stars. That’s not a throwaway number. For a solo-built macOS utility with no marketing budget and an MIT license, that’s a real signal that people actually wanted this.

Let me back up for a second and explain why this category matters right now, because it’s easy to shrug at “local AI” tools and assume they’re either performance-limited or built by people who went too far down the self-hosting rabbit hole.

The practical case for on-device transcription is stronger than it’s been at any point before 2026. Apple Silicon chips got genuinely fast at running quantized models. Whisper-class speech recognition at good accuracy no longer requires shipping audio to a third-party API. And the privacy angle isn’t paranoid posturing. Enterprise users, lawyers, doctors, journalists, and anyone doing sensitive interviews has a legitimate reason to not want meeting audio sitting on somebody else’s server. The market for “transcription that actually stays private” is not niche. It’s just that most products in this category are still betting on the cloud because the cloud is easier to build on and easier to charge recurring fees for.

Ghost Pepper bets the other way. It’s free, it’s open source, and the project page on Product Hunt lists support for 50 or more languages, which is not a small detail. Multilingual speech recognition at local inference speeds on consumer hardware is the kind of thing that would have required serious hardware two or three years ago.

The workflow integration piece is where I think it actually earns its keep in daily use. The Control-to-talk shortcut works everywhere, which sounds trivial but is the part that breaks in most competing tools. You know the experience: you install a dictation thing, it works great in its own window, and then you discover it has no idea how to paste into Slack or your terminal or your notes app. Ghost Pepper handles that through a global shortcut that just types into whatever is focused. No app switching. No clipboard drama. That’s the right call.

Meeting transcription saving to local markdown is also a very specific and deliberate choice. Markdown means the output is readable in any text editor, searchable with grep, version-controllable with git, and not locked into any proprietary format. It’s the kind of decision that signals the builder has opinions about software. Not everyone wants that (plenty of people would rather have a nice GUI to browse past meetings), but for developers and writers who already live in markdown, it’s genuinely useful.

I asked Matt Hartman about the decision to go fully local rather than offer a hybrid mode with optional cloud processing for users who might want better accuracy on noisier audio. “The whole point is that you never have to make that tradeoff,” he told HUGE. “Once you add a cloud option, you add a trust question. I wanted to remove the trust question entirely.”

That’s a principled stance and it carries a real cost. On-device models have limits. Background noise handling, heavily accented speech, and very long recordings will push the edges of what local inference can do well right now. The flip side is that you don’t have to worry about an API key rotating, a pricing tier changing, or your audio getting used for training data without your knowledge. Those are real concerns that a lot of people have quietly started caring about.

The Electronic Frontier Foundation has been documenting the ways that cloud-connected productivity tools can expose sensitive conversation data, and the concern has moved from “tin foil hat” to “standard enterprise compliance consideration” pretty fast. Ghost Pepper is positioning ahead of that shift rather than reacting to it.

On the open source side, MIT license means you can read the code, fork it, audit the model integration, and satisfy yourself that it’s actually doing what it says. That’s not a given even in tools that claim to be privacy-focused. There’s a difference between “we promise not to upload your data” and “here’s the code, check for yourself.” Ghost Pepper is doing the latter, and the 2,100 GitHub stars suggest people appreciate the transparency.

No account required. This is worth stating plainly because it’s increasingly rare. You download the DMG, drag the app to Applications, and that’s it. No email address. No license key. No onboarding flow. It just runs.

The distribution model is worth noticing too. Ghost Pepper ships as a DMG from GitHub releases rather than through the Mac App Store. That has tradeoffs. No App Store means no sandboxing requirements that would actually complicate local model access, no 30% cut, and no Apple review delays. It also means no automatic updates through the App Store mechanism and a slightly higher friction install for non-technical users. For the audience this tool is clearly aimed at (developers, technically literate knowledge workers, people who have strong opinions about where their data goes), a DMG from GitHub is fine. It’s probably the right call for this particular user base.

Product Hunt launch got solid traction and landed at rank 8 for the day. The comments were mostly from people who had already downloaded and tried it, which is usually a better sign than comments from people who just think the concept sounds cool.

The testimonials on the site include Ryan Hoover and Dave Morin, which is notable. Those aren’t paid placements (the site has no monetization and the MIT license makes the whole product free), so they’re organic endorsements from people who actually used the thing.

Ghost Pepper right now is a focused, well-executed tool for a specific use case: people who need reliable transcription, care about where their data goes, and are running Apple Silicon Macs. It does not try to be a full productivity suite. It doesn’t have a web app or an iOS companion or a team collaboration layer. Whether that expands is genuinely unclear, but the foundation, 2,100 stars, working software, no cloud dependency, and a clean open source license, is more than most solo-built tools have at this stage.