Does it work in VS Code, JetBrains IDEs, Cursor and Terminal?

Yes. Parakeety pastes the transcript by synthesizing a Cmd+V keystroke into whichever app has focus. That works in VS Code, every JetBrains IDE, Cursor, Zed, the system Terminal, iTerm, Warp, Ghostty, anywhere a text field accepts paste. The integration is at the operating-system level rather than at the editor level, which means there are no editor extensions to install or maintain.

Will it pick up code identifiers and proper nouns correctly?

Mostly yes for common identifiers like camelCase function names, snake_case variables, kebab-case package names, and standard stack vocabulary. Less reliably for unusual project-specific names. The model is general-purpose, not domain-tuned for code, so transcribing useEffect or kubectl apply works fine; transcribing the codename of a product nobody outside your team has heard of is luck-of-the-draw. Parakeety does not currently support custom vocabulary; if your work is dictation-heavy on a specific codebase with unusual identifiers, the 7-day free trial is the right way to find out whether the accuracy is enough.

Is it safe to use on code under NDA?

Audio does not leave the Mac. Transcripts do not leave the Mac. Both are processed on the Apple Neural Engine and the audio buffer is discarded immediately after transcription. The only outbound calls Parakeety makes are the one-time speech-model download from Hugging Face on first launch and periodic license checks against the license backend, which sends only the license key, a hash of a hardware ID, and the machine hostname. No audio, no transcripts, no analytics, no telemetry. For most NDAs that is a clean answer; whether your specific NDA's terms align is between you and your counsel.

How does it compare to cloud dictation that routes to LLMs?

Different shape of tool. Cloud dictation that routes to OpenAI or Anthropic for cleanup gives you polished prose at the cost of audio and transcript leaving your Mac. Parakeety gives you raw transcript at the cursor without anything leaving your Mac. For dictating a polished email, cloud is more direct. For dictating a commit message into git, a code-review comment into GitHub, or a quick reply into Slack, raw-at-the-cursor is what you actually want.

← Resources

Parakeety for engineers and developers

Most cloud dictation tools position themselves at writers and executives. They route audio to a server, transcribe it, optionally hand the transcript to an LLM for cleanup, and paste the result back. That is a reasonable shape for someone dictating an email, but it is not quite right for an engineer's day. Engineers spend half the day typing into different windows: Terminal in one space, an IDE in another, a GitHub PR in the browser, Slack DMs, Linear ticket descriptions, design docs in Notion. The bet that matters is not "better LLM cleanup", it is "paste at the cursor wherever the cursor happens to be, fast". That is what Parakeety does: local Mac dictation that pastes at the cursor.

Where dictation fits an engineer's day

The natural moments are the typing-heavy ones that do not need the visual focus of writing code:

Commit messages. Hold §, talk through what changed and why, release. The message lands in your editor or in git commit's scratch buffer.
Code review comments. Picking apart someone's diff in GitHub or Gerrit is the kind of task where talking through the logic is faster than typing it.
Design docs and ADRs. First-pass paragraphs in Notion, Confluence, or a Markdown file in the repo.
Slack and Linear. Replying to a thread, drafting a ticket, summarizing a debugging session.
Customer-facing replies. Support threads, incident comms, status-page notes, anywhere the words matter and the keyboard is in the way.

The thread connecting these is that they are all "cursor in a text field, dictate, move on". Parakeety pastes the transcript wherever the cursor is, whether that field is in Terminal, an IDE, a browser, or a native Mac app. It synthesizes a Cmd+V keystroke into whichever app has focus. Local dictation for writers covers the same first-draft case for prose work, in novels, journalism and academic papers.

Why paste-anywhere beats richer cloud features

Cloud dictation tools have richer features because they can: a server can run a bigger model, a richer prompt, a chained LLM cleanup. Parakeety has none of that. The trade is two-way:

You give up: AI cleanup, cloud-side custom vocabularies, command-mode editing, voice macros.
You get: an answer that fits everywhere your cursor goes, runs at 3,333x the speed of audio on Apple Silicon, and does not depend on your wifi or someone's API quota.

For the engineering use case the second column tends to win. The places engineers dictate are not the places that benefit most from LLM cleanup. A commit message you are about to publish does not need a model rewriting it. A Slack reply you are about to send should not have an LLM softening the punchline. The raw transcript at the cursor is the right shape.

Privacy under NDA

The harder constraint for working engineers is that the things you would dictate are often things your employer would rather not see hit a third-party server. Internal code names. Unreleased product names. Customer names in support threads. Architecture details for systems that are not shipped yet.

Cloud dictation handles this with policy. They describe what their servers do with the audio, you read it, you decide whether it is compatible with your employer's NDA or your customer's contract. Parakeety handles it differently: there is no server. Audio is captured to a memory buffer on your Mac, run through the model on the Apple Neural Engine, the transcript pastes, the buffer is discarded. There is no transmission to negotiate around because there is no transmission.

That is also the answer to the BYOK question. SuperWhisper and a few other apps offer "bring your own key" integration with OpenAI or Anthropic for post-processing, which is fine but means the transcript text leaves your machine to go through a model on someone else's servers. Parakeety does not do BYOK because there is no place in the pipeline to bolt cloud post-processing onto. We covered that comparison in Parakeety vs SuperWhisper.

What Parakeety is not

Worth being clear about. Parakeety is push-to-talk dictation with no AI cleanup, no command mode, no custom vocabulary, no voice macros. If your need is "dictate this loose audio into a polished email", Wispr Flow or SuperWhisper Pro are better tools for that. That is the workflow they were built for. If your need is "voice-control my IDE", Talon and Cursorless live in that lane and Parakeety does not compete with them. Parakeety is for the specific job of "cursor in a text field, talk, transcript pastes". The narrow scope is the point.

FAQ

Does it work in VS Code, JetBrains IDEs, Cursor and Terminal?: Yes. Parakeety pastes the transcript by synthesizing a Cmd+V keystroke into whichever app has focus. That works in VS Code, every JetBrains IDE, Cursor, Zed, the system Terminal, iTerm, Warp, Ghostty, anywhere a text field accepts paste. The integration is at the operating-system level rather than at the editor level, which means there are no editor extensions to install or maintain.
Will it pick up code identifiers and proper nouns correctly?: Mostly yes for common identifiers like camelCase function names, snake_case variables, kebab-case package names, and standard stack vocabulary. Less reliably for unusual project-specific names. The model is general-purpose, not domain-tuned for code, so transcribing useEffect or kubectl apply works fine; transcribing the codename of a product nobody outside your team has heard of is luck-of-the-draw. Parakeety does not currently support custom vocabulary; if your work is dictation-heavy on a specific codebase with unusual identifiers, the 7-day free trial is the right way to find out whether the accuracy is enough.
Is it safe to use on code under NDA?: Audio does not leave the Mac. Transcripts do not leave the Mac. Both are processed on the Apple Neural Engine and the audio buffer is discarded immediately after transcription. The only outbound calls Parakeety makes are the one-time speech-model download from Hugging Face on first launch and periodic license checks against the license backend, which sends only the license key, a hash of a hardware ID, and the machine hostname. No audio, no transcripts, no analytics, no telemetry. For most NDAs that is a clean answer; whether your specific NDA's terms align is between you and your counsel.
How does it compare to cloud dictation that routes to LLMs?: Different shape of tool. Cloud dictation that routes to OpenAI or Anthropic for cleanup gives you polished prose at the cost of audio and transcript leaving your Mac. Parakeety gives you raw transcript at the cursor without anything leaving your Mac. For dictating a polished email, cloud is more direct. For dictating a commit message into git, a code-review comment into GitHub, or a quick reply into Slack, raw-at-the-cursor is what you actually want.

Try it

Parakeety is a Mac menu-bar app. Hold the section key, talk, release; your words paste at the cursor in whichever app you were typing into. Audio never leaves the machine. There is a free 7-day trial with no card required. After that it is $30 once.

Try Parakeety free →