← Back to portfolio

Open-source macOS devtool

cmdFlow

A macOS app: under a global shortcut it runs clipboard text through an AI model (on-device Apple or OpenRouter/OpenAI). It also has vision — select a screen region and chat about it with the model.

Project preview

cmdFlow settings window with AI backend selection and actions.

Settings

AI backend (Apple / fallback / cloud), OpenRouter and OpenAI providers, and shortcut-bound actions.

cmdFlow actions list with global shortcuts and prompts.

Actions

Each action has its own global shortcut and prompt — e.g. translation and grammar fixing.

Overview

cmdFlow started from a simple annoyance: I keep copying a snippet and want to transform it fast — translate, fix grammar, summarize — without pasting into ChatGPT and switching windows. I press a shortcut, the clipboard text goes through my prompt into a model, and the result lands back in the clipboard.

Problem

Quick text processing usually means switching to a separate chat, pasting, waiting, and copying the result back. Across dozens of tiny operations a day, that is a real attention cost.

Product bet

If a text transformation is one global shortcut and the result lands straight in the clipboard, working with text becomes instant and never breaks focus.

What I built

  • Global shortcuts via Carbon, without Accessibility permissions.
  • Multiple actions, each with its own shortcut and prompt.
  • On-device backend: Apple Foundation Models (SystemLanguageModel).
  • Cloud providers OpenRouter and OpenAI on your own API key (Keychain), with a model search and web search.
  • Screenshot chat (vision): select a screen region and ask a model about it — streamed answers, multi-turn, chat history, editable system prompt.
  • Apple mode with automatic cloud fallback, launch at login, and an animated shortcut recorder.

AI layer

The core is the Apple Foundation Model running locally on the Mac. Testing surfaced a concrete limitation: the model rejects Polish input text (the unsupportedLanguageOrLocale guardrail). So I added cloud providers and an automatic fallback — private and on-device where possible, cloud where needed.

Architecture

Swift, SwiftUI (MenuBarExtra), FoundationModels, Carbon HotKey, NSPasteboard, Keychain, and SMAppService. An SPM package assembled into a .app by a script, with CI and releases via GitHub Actions on macos-26 (DMG + ZIP).

Key decisions

  • On-device and privacy as the default mode, cloud as a deliberate choice.
  • Cloud providers on the user’s key instead of my own backend and costs.
  • No .xcodeproj — an SPM package assembled into a .app, to keep the repo clean for open source.
  • Honestly document the Polish limitation instead of pretending it is not there.

Outcome / evidence

A public open-source project with GitHub releases (DMG/ZIP), a Vercel landing page, and three interchangeable AI backends.

What I would do next

A v0 landing page, a starter set of actions, and possibly signing and notarization once an Apple Developer account is available.

Contact

Tell me what you want to build.

Two sentences about the idea are enough. I’ll write back with where I would start, what I would skip, and whether I’m even the right person for it.

Email me