Skip to content

FAQ.md

Latest commit

 

History

History
77 lines (45 loc) · 4.73 KB

FAQ.md

File metadata and controls

77 lines (45 loc) · 4.73 KB
Raw

FAQ

Why not just use Playwright?

You should, when you own the site. Playwright is faster, has a stable API, and is great for CI. This project is for the other case: when the site actively resists automation. Playwright launches Chrome with CDP attached and exposes navigator.webdriver=true, window.cdc_*, and dozens of other fingerprintable signals. That makes it a one-line fingerprint for any modern anti-bot stack (Cloudflare Bot Fight Mode, Kasada, Akamai, DataDome). This project launches stock Chrome and drives it via the X server — the browser is indistinguishable from one a human is using.

Does it really bypass anti-bot?

For "nuisance" tiers (basic bot-detection scripts, Cloudflare Bot Fight Mode, reCAPTCHA v3 heuristics): yes, in most cases, because the stack emits no abnormal fingerprints. For "aggressive" tiers (hCaptcha enterprise, Kasada, residential-IP reputation checks on places like Ticketmaster): not alone. IP reputation dominates there, and no client-side stack fixes that — you need a residential proxy.

Why is it macOS / native-Linux hostile?

The core Python (server.py) works anywhere Python + X11 + xdotool exist. The scripts, systemd units, and docs are tuned for WSL2's quirks (wslrelay, WSLg, interop PATH, VM teardown). Supporting every Linux + macOS doubles the maintenance. If you need macOS, look at Anthropic's Claude Desktop computer-use or Codex Computer Use — they use native macOS APIs.

Why --no-sandbox?

Chrome's sandbox needs either CLONE_NEWUSER privileges or CAP_SYS_ADMIN to bootstrap its setuid/namespace sandbox. Inside WSL2 that's fragile. --no-sandbox is the documented workaround for sandboxed-container/WSL environments. The agent stack itself is your outer sandbox — run it in a disposable WSL distro (see WSL_SETUP.md).

Can I run multiple browsers in parallel?

Yes. Give each its own display number and VNC port:

CU_DISPLAY=100 CU_VNC_PORT=5901 bash scripts/display.sh start
CU_DISPLAY=101 CU_VNC_PORT=5902 bash scripts/display.sh start

Each MCP server pointed at a different CU_DISPLAY will see its own Chrome.

How accurate is the vision grounding?

Depends entirely on the model's vision capability. Large multimodal models (GPT-5 class, Claude Sonnet 4.6+) typically pinpoint buttons within ~5 pixels. Small inputs like radio buttons or 12px icons get missed ~10–20% of the time. The screenshot → act → screenshot → verify loop in prompts makes this self-correcting.

Can I use a different model / client?

Yes. Anything that speaks MCP over stdio works: Claude Desktop, Claude Code, Codex, mcp-inspector, custom runners. The examples reference hermes-agent because that's the test bed, not because of any coupling.

Why ship run_shell if it's a liability?

Debugging. An agent that can't ls /tmp/hermes-computer-use/chrome.log when something misbehaves is nearly useless to iterate on. The convention is: remove run_shell before production deployment. See SECURITY.md.

Will the ~5 min Chrome profile lock cause problems in parallel runs?

Yes, Chrome holds an exclusive lock on $CU_PROFILE_DIR. Use separate profiles per display:

CU_PROFILE_DIR=/tmp/hcu/profile-100 CU_DISPLAY=100 bash scripts/display.sh start

Does the Chrome profile save passwords?

Yes, by default. Do not store real credentials in CU_PROFILE_DIR. Either use a throwaway account, or scrub the profile between runs:

rm -rf "$CU_PROFILE_DIR"
systemctl --user restart computer-use.service

How do I see what the agent is doing, live?

Open http://localhost:6080/vnc.html in any browser while the agent runs. It's a real-time canvas of the Xvfb display.

Can I record automation runs to video?

Not built in. Run ffmpeg -y -video_size 1440x900 -f x11grab -i :99 out.mp4 in a side process. PRs welcome if someone wants a recording MCP tool.

What happens if Chrome crashes mid-run?

The agent's next screenshot will return a blank frame. Chrome does not auto-restart by default. Wrap it in a supervisor or use systemctl --user restart computer-use.service before the next attempt.

Is this affected by WSL's "mirrored" networking mode?

It works in both NAT and mirrored modes. Mirrored tends to be friendlier for reaching localhost:6080 from Windows without quirks. Both supported.

Do I need a GPU?

No. Xvfb is purely software. Canvas-heavy sites feel slow but function.

Why is the project called "hermes-computer-use" if it's agent-agnostic?

It was built alongside hermes-agent and that's the primary tested client. The name stuck. Any MCP client works.