I built a free system-wide AI writing assistant for Windows, macOS and Linux

#opensource #productivity #showdev #ai

Every day, writing with AI often means tedious context switching: highlight text, copy it, switch to a browser tab, paste into a chat UI, wait, copy the result, and paste it back over your draft.

Apple's writing tools showed a cleaner interaction: trigger AI directly where you type with a shortcut. But many of us work on Windows or Linux, or want to pick the model ourselves, such as Claude, Gemini, or an offline local model.

That led me to build AI Shortcuts (GitHub / Website), a free, open-source desktop app for Windows, macOS, and Linux that runs AI tasks directly on selected text across any application.

How it works

The workflow stays out of your way until invoked:

Select text in any app (browser, code editor, email client, notes).
Press your hotkey (Ctrl+Space on Windows and Linux, Option+Space on macOS).
Pick a command from the popup next to your cursor, or enter a custom prompt.
Receive the output: text replaces your selection directly (undo with Ctrl+Z / Cmd+Z) or opens in a response window where you can keep chatting.

If no text is selected when you press the shortcut, AI Shortcuts launches in chat mode, letting you ask questions from anywhere on your desktop.

Built-in commands and customization

AI Shortcuts provides eight built-in commands:

Proofread: corrects grammar, spelling, and punctuation while preserving meaning.
Rewrite: rephrases sentences for clearer expression.
Friendly / Professional / Concise: adjusts tone or eliminates filler.
Summary / Key Points / Table: condenses articles, notes, or transcripts into structured summaries.

With the Command Manager, you can create custom prompts, reorder actions, assign dedicated shortcuts with conflict detection, and import or export configurations. You can also paste images (Ctrl+V / Cmd+V) or attach text files directly into the prompt for multimodal context.

A real example: Proofreading in action

Here is a real example from the project website guide:

Hi Anna, i wanted to follow up on the meeting we had yesterday. Their were a few things we didnt finish, and I think we should of scheduled more time. Can you send me the slides befor friday so I can review them?

Selecting this draft and triggering Proofread returns the corrected draft alongside an itemized list of fixes:

Hi Anna, I wanted to follow up on the meeting we had yesterday. There were a few things we didn't finish, and I think we should have scheduled more time. Can you send me the slides before Friday so I can review them?

Corrections made:
- "i" → "I"
- "Their" → "There"
- "didnt" → "didn't"
- "should of" → "should have"
- "befor" → "before"
- "friday" → "Friday"

The response window shows the corrected text and the list of fixes:

Bring your own model (Cloud or offline Ollama)

AI Shortcuts works with your preferred model without recurring subscriptions:

Cloud providers: Google Gemini (free tier), Anthropic Claude, OpenAI-compatible APIs, Mistral AI, and OpenRouter (100+ models).
Local & offline models: Ollama, LM Studio, and llama.cpp. On Apple Silicon, on-device MLX models run natively.

You can set custom providers per command. For instance, route quick proofreads to a lightweight local Ollama model to stay offline, and reserve a larger cloud model for complex rewrites.

Privacy and key storage

Because desktop writing tools interact directly with your text, privacy is essential:

No telemetry: zero analytics, trackers, or ads.
Direct requests: your text goes only to your chosen AI provider when you trigger a command.
Key storage facts: on macOS, API keys are stored in the system Keychain. On Windows and Linux, provider keys are stored in the local settings file with light obfuscation (not encryption); per-command overrides use the operating system keyring when available. For complete offline privacy, pair the tool with local Ollama models.

Limitations and platform details

To be transparent about current constraints:

Windows: pre-built binaries are not code-signed, so Windows SmartScreen may show a one-time prompt ("More info" → "Run anyway"). Binaries include SHA-256 checksums and GitHub build provenance attestations (gh attestation verify).
Linux: requires an X11 session and xclip or xsel. Global hotkeys do not work under Wayland.
macOS: requires macOS 14 (Sonoma) or newer, and needs Accessibility and Screen Recording permissions in System Settings to capture selected text.