macOS menu-bar app. Select text in any application, press a shortcut, and Claudio rewrites it through Claude (or a local Ollama model), streams the result in a floating panel, then pastes it back in place of the selection.
Bring your own API key. Nothing goes through a third-party server: the app talks to the Anthropic API (or your Ollama instance) directly.
Download: claudio.okonoma.com or the latest release. Signed and notarized, macOS 14+, Apple Silicon. Unzip into /Applications.
Usage
| Shortcut | Action |
|---|
| ⌃⌥⌘K | Action palette: type to filter, press 1–9 or ⏎ to run. Anything that matches nothing becomes a free instruction. |
| ⌃⌥⌘I | Fix spelling, grammar and phrasing, keeping language and tone |
| ⌃⌥⌘P | Turn a rough idea into a clear prompt |
| ⌃⌥⌘^ | Turn it into a full prompt (role, context, task, constraints, output format) |
| ⌃⌥⌘F | Translate to French |
| ⌃⌥⌘E | Translate to English |
| ⌃⌥⌘T | Rewrite in a professional tone |
| ⌃⌥⌘R | Summarize |
| ⌃⌥⌘L | Explain in plain words |
| ⌃⌥⌘D | Free instruction ("translate to Spanish", "make it bullet points"…). Hold it to say the instruction instead of typing it. With nothing selected it becomes a request ("write a message to move tomorrow's meeting") answered at the cursor; selection or not, the track playing goes along, so it can be asked about |
| ⌃⌥⌘Space | Dictation: hold, speak, release — the text lands at the cursor. Tap it to go hands-free. A right-hand modifier held alone, such as right ⌥, can replace it |
The panel opens near the pointer and streams the result. ⏎ pastes it over the selection, or at the cursor when nothing was selected, and restores the clipboard; Esc cancels; ⌘C copies only. The source app keeps focus throughout. Shortcuts and system prompts are editable in Settings.
Dictation takes no selection: hold ⌃⌥⌘Space, speak, release, and the text is pasted where the cursor was — or tap it, and it listens hands-free until the next press. Recognition is Apple's own and runs on the Mac; the chosen model then punctuates the transcript and drops the hesitations, unless that model is Raw, which pastes it as it was heard. Whatever is playing pauses while it listens, so music doesn't talk over the voice, and resumes once the microphone closes. A vocabulary in Settings keeps names and jargon spelled your way (heard → written replacements, plus the terms passed to recognition and cleanup). A second shortcut listens in another language, each shortcut picks what its dictation becomes — cleaned up, translated to English, or turned into a prompt, all in that same single call — and the last 50 dictations are listed in Settings, text only.
Two things on first launch: an Anthropic API key (stored in the Keychain; ANTHROPIC_API_KEY overrides it in development) and the Accessibility permission, which macOS requests on the first shortcut. Dictation adds two more, asked on the first press: Microphone and Speech Recognition.
Build from source
Scripts/build_app.sh
open /Applications/Claudio.app
Swift Package Manager, Xcode 16+. The script builds in release, signs, and installs the bundle; signing and notarization options are documented in its header.
How it works
- Capture (Selection/): reads the selection through the Accessibility API, falls back to a simulated ⌘C for Chrome and Electron apps, and snapshots the clipboard to restore it after pasting, images and RTF included.
- Actions (AI/ClaudioAction.swift): one enum case per action, carrying its system prompt and token budget. What is actually sent is a ClaudioRequest, built from a catalog entry or a free instruction. The selected text travels inside a
<texte_source> tag so that a selection like "summarize my emails" is transformed, not obeyed. With nothing selected, a free instruction goes out as a request under a prompt of its own (AI/FreeRequest.swift), inside a <consigne> tag. Whatever is playing follows a free instruction in a <morceau_en_cours> block, read alongside the capture and waited for only when the request leaves, and the panel names the track under the answer; catalog actions never receive it.
- Providers (AI/AnthropicClient.swift, AI/OllamaClient.swift): SSE streaming, shared
TextStreamClient protocol. Parsing is separated from transport so it is tested on recorded transcripts. Model is chosen per action: Haiku 4.5 by default, Sonnet 5, Opus 5, or any local Ollama model.
- Panel (UI/ResultPanelView.swift): a floating window that never steals focus. Stream fragments are batched every 60 ms; the window height follows the content.
- Galette (Galette/): with the Galette music app installed, a track named in the panel gets Artist and Album buttons that open it there by name, through links whose values are percent-encoded byte by byte. A browser's track without an album names a channel rather than an artist, so it only gets the artist of an "Artist – Title" title.
Tests
swift test # unit tests, seconds
Scripts/test.sh --smoke # + every screen rendered to PNG, no network
Scripts/test.sh --release # + real API calls
CI runs the first two levels on every push and pull request and publishes the rendered screens as artifacts. TESTING.md describes what each level protects.
Known limits
- Simulated ⌘C/⌘V use physical key codes: fine on QWERTY and AZERTY, not on Dvorak or Bépo.
- "Launch at login" requires the packaged app in /Applications, not
swift run.
License
MIT. Independent project, not affiliated with Anthropic. Claude is a trademark of Anthropic, PBC.
