Hermes Agent had a monster day. The repo crossed 200,000 GitHub stars, and the computer_use tool went cross-platform. Same day. Both are big deals on their own. Together they tell you something is shifting.
I have been tracking this project closely, and I covered the v0.17.0 release just a few days ago. But this is a different kind of milestone. Let me break down both.
The 200K Star Signal
Nous Research posted the milestone on X: 200,000 GitHub stars for Hermes Agent. The GitHub API showed 199,726 when I pulled the numbers, which means it is fluctuating right around that line. Star counts bounce day to day, but the trajectory is clear.
For context, the repo also has 35,497 forks. That is not a vanity metric. People are cloning this thing, modifying it, building on top of it. The watchers count sits at 787.
After watching this project ship async subagents, iMessage support via Photon Spectrum, and now cross-platform desktop control in the span of two weeks, I believe the description.
Hermes Agent is not standing still. Check the project homepage if you want the full picture.
Computer Use Goes Cross-Platform
The bigger news for builders: Hermes Agent’s computer_use tool now drives cua-driver on macOS, Windows, and Linux. Before today, it was macOS only.
PR #50552 merged this morning. It was authored by Francesco Bonacci, the cua-driver creator, and merged by Teknium. The diff is 4,230 additions across 22 files. This is not a tiny patch.
It supersedes PRs #44221, #50507, and #30660, plus several duplicate attempts that tried to crack the Windows and Linux problems separately.
Here is what changed:
The runtime now detects your OS and dispatches the right installer. Architecture detection covers darwin-arm64, win-amd64, win-arm64, linux-x86_64, and linux-aarch64.
Eight cua-driver decoupling surfaces are consumed, each degrading gracefully on older drivers. That means if you are running an older cua-driver version, things still work. You just do not get the new structured features.
There is a new hermes computer-use doctor command that gives you a structured health check per platform. Exit codes tell you if things are ok, degraded, or blocked. On Windows, UTF-8 is forced on subprocess pipes to avoid encoding crashes.
The PR body includes a full validation table: 216 tests passing across computer_use, doctor, install, and capture routing suites.
The skill moved from skills/apple/macos-computer-use/ to skills/computer-use/ with platforms: [macos, windows, linux]. The interface is identical. If you were already using Computer Use on macOS, nothing changes in your workflow. Your existing prompts, your existing scripts, your existing automation, all of it keeps working.
What Was macOS-Only Before
To understand why this matters, you need to know what Computer Use actually does. It is Hermes Agent’s desktop automation layer.
The agent can capture your screen, identify UI elements by index, click buttons, type text, scroll, drag, take screenshots, and launch or kill applications. All of this happens in the background, without stealing your cursor or keyboard focus.
Before this PR, all of that only worked on macOS. The underlying driver, cua-driver, used macOS-specific APIs: Accessibility framework for element discovery, CGEvent for input simulation, Quartz for screen capture. If you were on Windows or Linux, the computer_use tool simply was not available to you.
Now cua-driver-rs handles the platform abstraction. On Windows, it uses UIA (UI Automation) for element discovery and SendInput for keyboard/mouse simulation. On Linux, it uses AT-SPI for accessibility and X11 for display interaction, with Wayland support via XWayland. The same Hermes Agent interface sits on top of all three.
Why This Matters for Builders
The key line from my post: “Same computer_use tool. Same background desktop control.”
That is the whole point. Hermes Agent does not give you a different tool for each OS. You get the same capture, click, type, scroll, screenshot interface regardless of whether you are on a MacBook, a Windows desktop, or a Linux box. The underlying driver changes. The API does not.
For builders, this means you can write one agent workflow and run it anywhere. No platform-specific code paths. No “works on my Mac” exceptions. If you are building an AI agent that needs to interact with desktop applications, you now have a single cross-platform abstraction that actually works.
The PR tested this across three environments: macOS (native), Windows 11 24H2 (VM), and Linux 24.04 (VM with Xvfb). Each platform went through install, doctor, status check, focused capture, click-by-index, type, screenshot delivery, and end-to-end chat with computer_use tool calls. All three passed.
The Infrastructure Play
Separately, these are both good stories. Cross-platform Computer Use is a meaningful capability upgrade. 200K stars is a meaningful popularity milestone.
Together, they signal something else: Hermes Agent is crossing from “interesting open-source project” to “serious agent infrastructure.” Popularity without capability is hype. Capability without popularity is a niche tool. Both at the same time, on the same day, is the inflection point.
I have been saying this since my original Hermes Agent review, but the project keeps proving it faster than I can write about it.
The v0.17.0 coverage from a few days ago already feels like a snapshot of a different era.
What to Watch
The follow-up PRs are already landing. Teknium opened #50842 to disable cua-driver telemetry by default, switching to an opt-in model. f-trycua opened #50855 to fix a broken pre-install asset probe that was false-negating on Linux, Windows, and Intel macOS. The community is iterating fast on the cross-platform foundation.
If you have been waiting to try Hermes Agent’s Computer Use on Windows or Linux, now is the time. The same tool, the same interface, three operating systems. That is the kind of change that matters.
