Linda is a free, private AI colleague designed specifically for Mac. It runs on local AI models and is designed to help users complete small, frequent daily tasks that have no ready-made integration coverage, such as getting invoices from three portals, comparing prices from five stores, naming a batch of scanned files, and so on. Users only need to describe the need in one sentence, and Linda will plan and execute the task, and before performing irreversible operations such as sending, purchasing, deleting, publishing, or submitting personal data, it will show an approval card and wait for the user's confirmation.
In terms of main capabilities, Linda has an independent browser profile, can log in, click, read, and download in the background, and will not take over the user's mouse or open the user's own account; users can watch in real time through the Computer column, take over at any time, or close the window and continue working. For questions that only require reading, Linda can search and read directly without opening a window. Linda supports voice interaction; hold the right Option key to speak like Siri, speech recognition runs on-device, and audio does not leave the Mac. Each task has an independent folder, and tasks under a project share the same folder, making it convenient for files to continue being used; Linda can read, rename, and merge files there, and generate PDFs, spreadsheets, and documents, while other areas remain inaccessible. After a task is completed, it can be turned into a daily, weekly, or monthly routine with one click, and pages can also be checked on a schedule. Users can also create "experts" for repetitive work, such as Researcher or Bookkeeper. Each expert has a custom name, role, instructions, skills, starting project, and toggles for browser, files, commands, and memory, but turning off toggles only reduces tools, and the approval step always remains.
In terms of usage, Linda uses parallel search by default, with no key required, and users can also choose Brave, Tavily, or Exa and add their own key, with DuckDuckGo always available as an option and fallback. During search, only the search terms, a short line in the request, and a random ID are sent; memory, files, or conversations are not sent. Linda stores personal information told by the user locally in "Settings > Memory," which can be read, edited, and forgotten at any time. Text in web pages, emails, or files will not be saved as facts about the user. Linda can connect to Outlook, Teams, Numbers, Pages, Notes on the Mac, as well as Calendar, Reminders, Notion, Todoist, Linear, GitHub, Stripe, and user-writtenskills, and all apps require approval.
In terms of AI models, Linda can run completely locally, using MLX on Apple Silicon to run assistants from 2B to Qwen3 Next 80B on its own, for example, a 4B assistant on M2 Pro at about 60 tokens per second; it can also connect to an existing Ollama, LM Studio, or mlx-lm, or use its own key for frontier models such as OpenRouter, OpenAI, Anthropic, Google, Groq, Mistral, and DeepSeek. No matter which is chosen, memory, voice, files, and passwords remain on the Mac.
Applicable scenarios include handling invoices, comparing prices, organizing scanned files, regularly checking web pages, completing repetitive research or bookkeeping, and so on. In terms of limitations, accessing websites or searching requires the internet; an Apple Silicon Mac (M1 to M5) and macOS 26 are required; after installation, two permissions need to be granted: microphone and accessibility; download coming soon; free to use, and if using frontier models, you need to bring your own key and pay according to usage.
