Offloop

Offloop is a dictation app for the Mac that turns spoken thinking into clean written text without sending a single byte off the machine. It's built for developers and AI power users, people whose bottleneck usually isn't figuring out what to say but typing it out fast enough and spelling the technical terms correctly on the way through. The framing on the site is blunt about the use case, which is talking to your agent instead of typing to it, and that one decision shapes almost everything else in the product. Prompts to coding assistants have gotten long, and typing a paragraph of context to a model that reads it in a second has quietly become the slow part of the loop.

Activation is a single key. You hold it down while you talk, or tap it twice when you want a longer hands-free session, and the finished text lands wherever the cursor happens to be blinking. The site puts the round trip at under 200 milliseconds, which is the practical difference between dictation you keep using and dictation you quietly abandon after a week. Anything slower and you start waiting on the tool instead of thinking out loud, and the moment you notice the lag you go back to the keyboard. There's no window to open, no separate app to switch into, and nothing to click before you start talking.

What separates it from the dictation already built into macOS is the cleanup pass. Unscripted speech is full of restarts, filler, and sentences that change direction halfway through, and most transcription tools hand all of that back to you verbatim, which leaves you editing a wall of text you just finished speaking. Offloop polishes what you said into something readable, dropping stutters and repairing terminology as it goes, while holding to a rule the site states plainly, which is that it adds nothing you did not say. That constraint matters more than it sounds. A cleanup model willing to invent produces text you have to read twice before you trust it, and at that point you've given back the time you saved.

Screen Context is the feature that makes it usable for technical writing. With it turned on, Offloop reads one frame of what's currently visible on screen and uses that as a reference while it cleans up your speech. If a library name, a function, or a snake_case identifier is sitting right there in your editor, the transcript spells it the way your code spells it rather than guessing phonetically, and phonetic guessing is exactly where generic speech-to-text falls apart for programmers. It also learns the shorthand and library names a given project leans on and applies them automatically, so you stop correcting the same handful of words every single day, and code formatting conventions like backticks and snake_case survive the trip into the text field.

Privacy here is structural rather than a promise buried in a policy document. Speech recognition, the screen read, and the cleanup step are all one small model running locally on the Mac, so your speech, your screen, and your text never leave the device. Audio is held only while the key is down and is gone the moment you let go, and after the first model download the whole thing keeps working offline. That posture is worth more than it looks for anyone under a client agreement that forbids sending source code to a third-party server, or anyone who'd rather not stream a live view of their screen to a vendor in exchange for better autocorrect on library names.

It works anywhere a text cursor does, and the applications the site names tell you exactly who's expected to use it, including Claude, Terminal, ChatGPT, Windsurf, Pi, OpenCode, Herdr, and Conductor. That's a list of coding agents and shells rather than office software, so the tuning is aimed at someone dictating a long prompt to an agent or talking through a commit message, not a journalist transcribing interviews or a student writing an essay. The price of running everything locally is hardware. You'll need Apple Silicon, an M1 or newer, running macOS 14 Sonoma or later, and there's no Windows build, no Linux build, and no web version, because the model has to run on the chip in front of you.

Access is freemium and priced modestly. The free tier covers 2,000 words a week with no credit card, which is enough to work out whether talking instead of typing actually fits the way you work, and because everything runs on device the free tier is a real offline product rather than a crippled trial. Pro costs $9 a month for unlimited words plus priority support, and new signups currently get two months of Pro at no cost. For something you'd reach for dozens of times a day that's a low bar, and the flat monthly price means there's no per-minute meter ticking in the back of your mind while you're mid-sentence.