DeepSeek Harness computer use is real, but it's experimental, macOS-only for now, switched off by default and limited to PTC sessions. You turn it on in Plugins, grant two macOS permissions, and the agent can then take screenshots, read app windows, click, scroll and type on your Mac. Windows support is still in development, according to the project's own docs.
📺 Watch: NEW DeepSeek Harness v0.2 is INSANE!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
That's the short answer. Below is exactly how it works, how to switch it on, the second route through Cua Driver, and how it stacks up against the computer use built into Hermes Agent.
Can DeepSeek Harness control your computer?
Yes. DeepSeek added it in the 15 September 2026 release (v0.1.6-alpha.1), and the release notes say it plainly: experimental Computer Use "through Cua Driver MCP or the native driver for local computer interaction and screenshots." The same release added experimental Browser Use with three backends: Playwright MCP, Chrome DevTools MCP and Stagehand.
Here's what most people miss. The download page at deepseek.com/en/harness doesn't mention computer use at all. It talks about files, coding, research, background tasks and plugins. The feature lives in the GitHub repo, under the experimental packages, and you have to go looking for it.
Three limits to know before you start:
- macOS only. The Computer Use bundle needs a macOS host, and the native helper needs macOS 14 or later. Windows is in development.
- PTC sessions only. Computer use is only installed in PTC mode sessions. Standard, Minimal and Creator sessions don't get the tools.
- Source builds first. The bundle README says the workflow targets developers running the macOS Web UI or the unpackaged Desktop app from source. Distribution and signing for the packaged Desktop helper is listed as deferred work.
If you want help wiring agents like this into a real business, the AI Profit Boardroom ($69/mo) gives you my Agent OS, a 30-day AI agent roadmap, daily tutorials as new tools drop, 4 coaching calls a week and the prompt library. Over 3,400 members are in there.
How to turn on DeepSeek Harness computer use
This is the built-in route. DeepSeek calls the shared capability "computer use" and the bundle is listed in the app as Computer Use.
- Run Harness from source on a Mac. From the repo root, run pnpm run build, then pnpm dsh web. The Web UI opens at 127.0.0.1:3080.
- Open Plugins. Select More beside Experimental plugins, open Computer Use and flip the switch.
- Grant permissions. The plugin's own description says to switch to PTC mode and grant Accessibility and Screen Recording. The permission page tells you which app macOS sees as responsible, so you know what to enable in System Settings.
- Set PTC as your default. The detail page has a Set to PTC button for new sessions. Existing sessions keep their mode.
- Approve each app. The first time the agent picks an app, you get asked to allow it for the session, deny it, or save the access. Deny once and it stops asking for that session.
The switch is saved per profile, so enable it in every profile that needs it. Turn it off and the tools and their model guidance disappear.
Once it's on, the agent can list running apps, find installed apps, read accessibility trees, capture window screenshots, click, hover, scroll, drag, type and paste. The model guidance tells it to work in the background where it can, and to check fresh state after every action, because a delivered click doesn't prove the job got done.
Want me to look at whether a setup like this fits your workflow? Book a free AI strategy session and we'll map it out.
📺 Watch: DeepSeek Harness: FREE 1 Hour Course!
The Cua Driver route: computer use over MCP
The second route skips DeepSeek's own macOS helper and uses Cua's cua-driver instead. The computer-use docs list two Cua Driver providers, and both are experimental npm packages you have to activate yourself:
- Cua Driver MCP. Connects to a cua-driver executable you've already installed. The minimal config mounts @deepseek-ai/dsh-computer-use plus @deepseek-ai/dsh-experimental-computer-use-cua-driver-mcp, with command set to cua-driver and args set to mcp. The driver's tools show up under the mcp__cua-driver-mcp__ prefix.
- Cua Driver native. Runs the driver inside the Harness host through npm, so there's no separate CLI to install. The README warns that a native crash can take down the host process.
Only one computer-use provider can be loaded at a time. Load a second one and it fails. Installation and desktop permissions for the MCP route stay with Cua Driver's own setup guide.
This works because Harness has a proper MCP client. Per the MCP docs, you add one @deepseek-ai/dsh-mcp-client entry per server, over stdio for a local program or Streamable HTTP for a service. No server is on by default. So any MCP server you trust, not only Cua's, can plug in the same way.
📺 Watch: DeepSeek Harness VS Claude Code
What DeepSeek Harness does well without touching the screen
Most jobs don't need an agent clicking around your desktop. Harness is strongest when it works through files and the terminal, and that's where I'd start.
- Files. DeepSeek pitches it for organising files, analysing data, drafting documents and building slides, with previews for .docx, .xlsx, .html and .pdf output.
- Terminal. Terminal is a listed plugin, and the shell packages cover bash and PowerShell. Since v0.2.0-rc.2 on 29 September 2026, the dsh command ships bundled with the macOS and Windows Desktop apps.
- Scheduled tasks. The schedule package supports one-shot, fixed-rate, daily, weekly and cron reminders that survive restarts. Each one lands as a follow-up message in its original conversation, and you manage them from the Automation tasks page.
- Browser use. Separate from computer use. Playwright MCP, Chrome DevTools MCP or Stagehand can drive Chromium, and each one only switches on when you mount it.
For the full desktop app walkthrough, see my DeepSeek Harness Desktop guide.
Capability table: what's supported and how
| Capability | Supported? | How |
|---|---|---|
| Desktop control on macOS | Yes, experimental | Plugins, Experimental, Computer Use switch; PTC mode; Accessibility and Screen Recording |
| Desktop control on Windows | Not yet | In development, per the bundle README |
| Desktop control on Linux | Not native | The bundle needs a macOS host; a browser elsewhere can connect to a Mac host |
| Cua Driver over MCP | Yes, experimental | Mount the cua-driver MCP provider with command cua-driver, args mcp |
| Cua Driver native | Yes, experimental | Mount the native provider; runs the npm driver in-process |
| Browser control | Yes, experimental | Playwright MCP, Chrome DevTools MCP or Stagehand provider |
| Any MCP server | Yes | One dsh-mcp-client entry per server, stdio or Streamable HTTP |
| Files and documents | Yes | Built-in file tools and previews |
| Terminal commands | Yes | Terminal plugin and shell tools |
| Scheduled tasks | Yes | Schedule reminders and the Automation tasks page |
DeepSeek Harness vs Hermes Agent for computer use
I run Hermes daily inside my Agent OS, with a profile per model, so this is the comparison I care about.
Hermes ships computer use as a built-in toolset called computer_use. Per the Hermes docs, it's background desktop control through cua-driver: screenshots, click, drag, scroll, type, key presses, wait, list_apps and focus_app. It doesn't steal your cursor or keyboard focus. It works with any tool-capable model, and it runs on macOS, Windows and Linux. You need cua-driver on your PATH, which you install through hermes tools.
Here's how they line up:
- Platforms. Hermes covers macOS, Windows and Linux. Harness's native route is macOS only right now.
- Setup. Hermes is one toolset plus cua-driver. Harness needs PTC mode, a plugin switch and, for now, a source build.
- Driver. Both can sit on Cua's cua-driver. In Harness it's one of two provider options. In Hermes it's the recommended path.
- Consent. Harness asks you to approve each app before it reads or touches it. That's a nice touch for a developer preview.
If you want the Hermes side in depth, read my Hermes computer use skill breakdown and the Hermes Cua setup guide. If you'd rather have me help you pick between them for your own work, grab a free AI strategy session.
Safety before you hand over the mouse
DeepSeek is upfront here. The SAFETY notice says Harness hasn't had a security audit and shouldn't be treated as production-ready. It can run model-generated code and commands, load third-party plugins, and touch files and credentials you give it access to.
Their advice is simple. Run it with the least access it needs, prefer a disposable VM or container, keep backups, and review commands before they run. With computer use, I'd add one more: only approve the apps the task actually needs.
To see how DeepSeek's models perform inside agents, check Goldie Bench. For the bigger picture of running agents like this across your business, start with the Agent OS guide.
FAQ
Does DeepSeek Harness have computer use?
Yes, as an experimental feature. It arrived in the 15 September 2026 release and works on macOS through the Computer Use plugin, or through Cua Driver over MCP or as a native npm driver.
Why can't I see the computer use tools?
Check three things. The Computer Use switch is on in that profile, the session is in PTC mode, and macOS has granted Accessibility and Screen Recording to the app that launched Harness.
Does it work on Windows?
Not yet. The bundle README says Windows support is in development. Until then, the Windows Desktop app still handles files, terminal work and scheduled tasks.
Is it better than Hermes computer use?
For cross-platform desktop control today, Hermes is further along. It covers macOS, Windows and Linux out of the box. Harness wins on per-app consent prompts.
Want to put DeepSeek Harness computer use to work this week? Join the AI Profit Boardroom for the Agent OS and daily tutorials, or book a free AI strategy session and we'll plan your setup together.
