DeepSeek’s AI agent now installs like a normal app on Mac and Windows

DeepSeek has released desktop versions of DeepSeek Harness, its AI agent app. There’s one for Macs with Apple’s own M-series chips and one for 64-bit Windows PCs, and both are free downloads from DeepSeek’s Harness page. Linux users get it the old way, as a package from npm, the online library developers use to install JavaScript tools.

DeepSeek is the Chinese AI company behind the DeepSeek chatbot. Last month we covered its plan to run its models on at least 160,000 Huawei AI chips. In AI, a harness is the software wrapped around a model that lets it act on your computer: read and edit files, run commands and work through a task in several steps without you typing each one. So instead of telling you how to fix a spreadsheet, it opens the spreadsheet and fixes it. DeepSeek’s page lists organising files, analysing spreadsheets, drafting documents and slides, writing and testing code, researching with cited sources, and running jobs on a schedule, such as a weekly report every Friday.

DeepSeek open-sourced Harness in August under the MIT licence, which lets anyone use, change and resell the code. It’s built so that every part of it is a plugin: the model, the tools, the skills and even the interface can be swapped out, and you can ask the agent to write a new plugin for you. DeepSeek still labels it a developer preview and warns that updates will break compatibility with older versions.

What the desktop app changes

Until now, running Harness meant installing Node.js, a programming environment, and then typing npx @deepseek-ai/dsh web into a terminal, which opened Harness in your browser. That’s routine for a developer and a wall for almost everyone else.

The desktop app puts the same interface in its own window and carries everything it needs inside it: its own copies of Node.js and Python, plus the Python libraries that read and write Excel, Word and PowerPoint files. That’s why the downloads are big, about 370MB for the Mac version and 289MB for Windows, so mind your data bundle. In return, you install it like any other app and never touch a terminal.

DeepSeek doesn’t list a version for older Intel Macs. Owners of those can still run Harness through the npm route in a browser.

What it costs to use

On first launch, the app asks you to sign in with a DeepSeek account through your browser, or to paste in an API key, the password-like code that lets an app bill your DeepSeek account directly. Signed-in users can check their usage and top up credit from inside the app. Harness also takes keys from other AI companies, including OpenAI and Anthropic, if you’d rather use their models.

DeepSeek charges per token. Its pricing page defines a token as the smallest unit of text the model recognises, which can be a word, a number or a punctuation mark, and bills for what goes in and what comes out. The default model is DeepSeek-V4.1-Flash, which DeepSeek launched on 10th September 2026. Prices are in US dollars, converted here at KES 129.56 to the dollar, the rate on 7th October 2026.

“Cached” input is text the model has already seen in the same conversation. That’s most of what an agent sends, because it passes the whole conversation back to the model at every step, and the repeat text is billed at the cached rate.

Say the agent works through a folder of sales reports over 20 steps, and that adds up to 2 million input tokens, of which 1.5 million are cached and 500,000 are new, plus 100,000 tokens of output. At peak rates that’s USD 0.28, about KES 40. Off-peak, it’s about KES 20. A long coding job can run to many times that, so check the usage page after your first few tasks.

DeepSeek’s peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, which in Kenya is 4am to 7am and 9am to 1pm. Everything else is off-peak, including weekends and Chinese public holidays. So heavy jobs run after 1pm cost half what they would at 10am.

Safety and data collection

DeepSeek’s own safety notice says Harness hasn’t had a security audit and that mistakes by the model or a bad plugin can modify or delete files, or expose data and passwords on the machine. The app asks before actions that need approval, but DeepSeek says those prompts don’t guarantee protection. It recommends keeping backups, running Harness in a separate environment where you can, and reading commands before you allow them.

The desktop app also reports usage to DeepSeek by default, such as which pages you open and which model you pick, and there’s no switch for it in the app. DeepSeek’s developer documentation says prompts, responses and API keys are never part of those reports.

If you’ve wanted to try an AI agent on your own computer, Harness is now an ordinary install. But it’s preview software that can delete files, so give it a copy of a folder to work on until you’ve seen how it works.