TextWren

TextWren is a small macOS app for everyday writing. It translates, fixes grammar, and runs your own prompts on any text, either in its own window or right where you are typing.

• Popup in any app. Select text, press ⌘C twice, get the result, then Replace the selection or Copy it.
• Main window for longer text, with results that update as you type.
• Formatting is kept. Lists, links, bold, italic and code survive the round trip to the model and back, including when you paste into Slack, Notes or Google Docs.
• Up to five custom actions such as Grammar, Summarize or "Create task", each with its own model and reasoning level.
• Voice dictation. Hold a shortcut and speak into any app, with a local speech model (audio stays on your Mac) or OpenAI. A menu bar icon and popup show what is happening.
• Agents and live conversation (Beta). Ask remote agents (HTTP APIs or MCP servers) by voice and hear the answer, or talk live with GPT-Live-1 while it uses your tools.

Requirements

• macOS 14.6 or later (Apple Silicon recommended for the on-device models)
• An OpenAI or Google Gemini API key, or none at all: the on-device text and speech models run without a key

Install

1. Download TextWren.zip from the latest release and open it.
2. Move TextWren.app to Applications.
3. The app is ad hoc signed, not notarized. If macOS blocks the first launch, Control-click TextWren.app and choose Open, or use System Settings → Privacy & Security → Open Anyway.

TextWren was called TinyAI before version 2.1.0. Updating keeps your settings, API keys and downloaded models. After installing TextWren.app, delete the old TinyAI.app; if Accessibility does not work, remove the old entry in System Settings and add TextWren.app.

Build from source

Open TinyAI.xcodeproj in Xcode and run the TinyAI scheme (the app target keeps its original name). The build phase downloads the pinned transcribe.cpp and llama.cpp frameworks into Vendor/. To compile the release app from the command line without installing it, run:

zsh scripts/build_and_install.sh --build-only

Without --build-only the script signs the app, installs it into /Applications, replaces an older version and relaunches it. It needs one of the maintainer's signing identities, described below.

For the maintainer's Mac, the script uses the TinyAI Local Code Signing identity in the login Keychain. It pins that certificate's SHA-1 fingerprint in the script so a different certificate with the same name cannot silently sign an update. If that identity is unavailable, the script can use an Apple Development certificate for team Y29LYS5D8M. The script refuses to install a version that is not newer than the installed one. Bump MARKETING_VERSION and CURRENT_PROJECT_VERSION in Xcode and the matching version and build_number at the top of the script together. Use --build-only to compile without installing. For an ad hoc signed release archive without a certificate, run zsh scripts/sign_release_ad_hoc.sh /path/to/TextWren.app after building and before creating the ZIP. Ad hoc signatures can change between builds, so macOS may require Accessibility to be granted again after an update.

The local identity is for this Mac only. Keep its private key in the login Keychain, and back up the certificate and private key securely before migrating to another Mac. Creating a new local certificate changes the app's signing identity. Local signatures cannot be notarized or used for public distribution. macOS still requires each permission to be granted initially; stable signing only gives later updates a consistent code identity, and permission retention should be checked on the target Mac after updating. The local build uses a Library Validation runtime exception to load its bundled CTranscribe framework because a self-signed certificate has no Apple Team ID. The framework is still signed with the same local certificate and verified before installation. Apple Development builds use the standard entitlements without this exception.

Set up

1. Open Settings → API and paste an OpenAI and/or Gemini key. Keys are checked before they are saved. To stay fully offline, download the on-device models there instead.
2. In Settings → Primary, choose what the two result panels show: the built-in Translate or one of your actions.
3. When macOS asks, allow Accessibility for TextWren (System Settings → Privacy & Security). The popup hotkey, voice shortcuts and Replace need it. If a previously enabled entry does not work after an update, remove it, add /Applications/TextWren.app, enable it and relaunch.
4. For voice features, allow Microphone access when asked (or in Settings → Voice).

Settings are applied when you press Save; Cancel discards the changes.

Use

Popup

• Select text in any app and press ⌘C twice. The hotkey and single/double press mode can be changed in Settings → Hotkeys.
• Replace returns to the original app and pastes the result over the selection. It does not paste if that app is no longer in front.
• Copy puts the result on the clipboard with formatting.
• Run other actions with the buttons or ⌘1 … ⌘5.

Main window

• Paste or type into Source text. A new request starts shortly after you stop typing and replaces any request still running.
• Choose a target language, or Auto to translate between your main and additional languages (set in Settings → Primary).

Dictation

By default, hold fn⌃ (fn and Control together) and speak, or tap them to start and tap again to stop; Esc cancels. The text is pasted where the cursor was, and your clipboard is put back afterwards. If no text field has focus, the text appears in the popup under the menu bar icon with a Copy button. A shortcut can be modifiers alone, like fn⌃, or modifiers with a key, like ⌃V. Shortcuts with a key use the physical key, so they work in any keyboard layout (⌃V and ⌃М are the same shortcut). Existing saved shortcuts are preserved. Change the shortcut and engine in Settings → Voice:

Engine	Model	Good for
Local	Voxtral Mini 4B Realtime	Live text, 13 languages. Examples: English, Mandarin, Hindi, Spanish, Arabic. Model weights need about 2.8 GB RAM, plus runtime memory.
Local	Nemotron Streaming 3.5	Fast live text, 28 languages in TextWren. Examples: English, Mandarin, Hindi, Spanish, Arabic. Model weights need about 750 MB RAM, plus runtime memory.
Local	Canary 180M Flash	Small offline model. Its four languages are English, Spanish, French and German. Model weights need about 220 MB RAM, plus runtime memory.
OpenAI	GPT Live Transcribe (default)	Streams while you talk. OpenAI list price: $0.017/minute.
OpenAI	GPT Transcribe	Transcribes after recording stops. OpenAI list price: $0.0045/minute.

When a new installation has neither an OpenAI API key nor a downloaded local model, TextWren offers to download Nemotron for local dictation. Download and RAM figures are approximate; actual memory use depends on the runtime and recording length.

Local models are the ones Handy uses, run with transcribe.cpp on the Mac's GPU. Download them in Settings → Voice or Settings → API; each file is checked against a pinned SHA-256 and stored in ~/Library/Application Support/TinyAI/Models (the folder keeps the app's original name). Audio for local models never leaves the Mac.

Agents (Beta)

Agents and live conversation are a beta. Turn them on in Help → Agents and live conversation (Beta); this adds the Agents and Live tabs to Settings and their shortcuts to the menu bar.

Settings → Agents has two lists:

• Connections: an HTTP API (URL, method, JSON body template, and the path to the answer in the response) or an MCP server (Streamable HTTP). Authorization is either a secret header stored in the Keychain, or, for MCP servers, Sign in with browser: OAuth in your default browser with automatic client registration. Test & load tools lists the server's tools.
• Agents: each has its own voice shortcut and a connection (for MCP, the tool and a JSON arguments template). Templates can use {{transcript}}, {{sessionId}}, {{language}} and {{agent}}. Follow-ups within the configured minutes reuse the same {{sessionId}}.

Hold an agent's shortcut and ask. The answer can appear in a floating panel, be pasted at the cursor, be copied, and be spoken with an OpenAI voice (with a style prompt) or a macOS voice. Requests wait up to the connection's timeout, so agents can think before answering.

Live conversation (Beta)

Settings → Live sets a shortcut that starts and ends a full-duplex conversation with GPT-Live-1. It keeps talking while its backend model (for example GPT-6 Luna) reasons and calls tools: web search, the tools of the MCP servers you pick, and your agents. Tool calls run on the Mac with each connection's sign-in. The transcript can be shown in a panel.

Menu bar

The menu bar icon shows the state: ready, recording (red, pulsing), transcribing, an agent thinking, speaking, live, or an error (orange). While you record, a popup under the icon shows the microphone level and the text as it is recognised; errors appear there too.

The menu starts dictation (and, with the beta on, agents or a live conversation), cancels the current one, lists the five most recent transcripts (click one to copy it), shows the last agent answer and the model download progress.

Custom actions

Create up to five actions in Settings → Actions: a title, a prompt, a model and a reasoning level. The text you select is treated as input for the prompt: requests or questions inside it are processed, not answered.

• {{targetLanguage}} in a prompt is replaced with the language chosen in the menu.
• Settings → Style adds shared tone or terminology to the actions you pick.

Example prompts:

Action	Prompt
Summary	Summarize in 3 bullet points, under 60 words.
Polite email	Rewrite as a polite, concise email. Keep names, dates and action items.
Shorter	Make the text about half as long without losing meaning.

Models

TextWren supports a fixed set of models, each with the reasoning levels its provider accepts.

Provider	Models
OpenAI	GPT-6 Luna, GPT-6 Sol, GPT-6 Astra, GPT-5.6 Luna, GPT-5.6 Terra, GPT-5.6 Sol, GPT-5.5
Google Gemini	Gemini 3.8 Flash, Gemini 3.5 Flash-Lite, Gemini 3.1 Pro
On-device	Qwen3.5 4B (2.7 GB download, about 3 GB RAM while loaded)

The default is GPT-6 Luna with Low reasoning. In testing it was as accurate as higher levels for translation and grammar, and faster. Use Medium or High for explanations and harder tasks. Settings that refer to a model that is no longer supported switch to the default.

Qwen3.5 4B runs on this Mac with llama.cpp, needs no API key, and the text never leaves the device. Download it in Settings → API, then pick it for translation or any action. In a blind comparison with GPT-6 Luna (Low) it scored 4.7 vs 4.8 of 5 for grammar but 3.8 vs 4.9 for translation. Reasoning is Off by default (it scored best). Low, Medium and High turn thinking on with a growing budget and are much slower. Typical wait for a sentence or two on Apple Silicon:

Reasoning	Wait
Off	1.5–2.5 s
Low	about 30 s
Medium	about 2 min
High	about 6 min

The model loads on first use and is freed after five idle minutes.

Model output has em dashes (—) replaced with hyphens (-); code is left unchanged.

Privacy

• API keys, connection header values and OAuth tokens are stored in the macOS Keychain.
• The microphone records only while a voice shortcut is active or a live conversation runs. With a local model, audio stays on the Mac; with OpenAI, the recording goes to OpenAI; agent requests send only the transcript to your connection.
• Text is sent to the selected provider only when an action runs: when you open the popup, or after a short pause while editing in the main window.

Development

• Unit tests: xcodebuild test -scheme TinyAI-UnitTests -destination 'platform=macOS' (CI runs them and the --build-only release build on every PR: .github/workflows/tests.yml)
• Manual checks: open scripts/manual-test-kit.html in a browser. It has rich-text samples (lists, links, code, tables, a long text) with the expected result for each.
• Model catalog and request rules (reasoning, token budget): ModelCatalog and LLMRequestPolicy in TinyAI/TranslationService.swift.
• Local speech engine: scripts/fetch_transcribe_cpp.sh downloads the pinned transcribe.cpp framework into Vendor/ (the build script and the Xcode build phase run it). scripts/fetch_llama_cpp.sh does the same for llama.cpp, which runs the on-device text model (LocalLLMEngine in TinyAI/LocalLLM.swift); its tests in LocalLanguageModelTests run only when the model is downloaded. The local-engine tests in TinyAITests/VoiceFeatureTests.swift run only when the model files are downloaded.
• Voice logs: /usr/bin/log show --last 5m --predicate 'subsystem == "IT.TinyAI"' --style compact shows hotkeys, recording level, transcripts and failures.
• Voice flow: VoiceCoordinator (hotkeys → recording → transcription → dictation, agents, live), MCPClient, OAuthBrowserAuthorizer, HTTPAgentClient, and the GPT-Live wire format in LiveProtocol.
• HTML handling: RichTextHTMLParser and RichTextHTMLSanitizer in TinyAI/RichTextPayload.swift. TextWren parses HTML itself instead of using AppKit's importer, which crashes on macOS 27 for HTML with links.

License

TextWren is released under the MIT License. The bundled transcribe.cpp and llama.cpp frameworks and the downloaded models keep their own licenses.