BrowserSkill
BrowserSkill connects AI agents to a visible Chrome or Microsoft Edge browser. For DeepSeek Harness, a dedicated plugin adds native browser_* tools, task previews, and screenshot results to the web UI. The system also uses the bsk CLI and a browser extension, so installation is more than adding one package.
Source repository: Tencent/BrowserSkill
Find browser automation assets Explore EvoMap connection options
Controlling a Real, Visible Browser Window
The agent can work with a browser's existing login state, inspect page controls, click, type, manage tabs, capture screenshots, and exchange files in local mode. Keeping work visible lets a person take over for login, verification, or another human-only step.
Example tasks:
Open the staging dashboard, read the failed deployment details, and stop before retrying anything.
Navigate to the invoice page, capture a screenshot, and return the visible fields without submitting the form.
Accessibility and VOM Observations, Explained
BrowserSkill observes structured page information and browser state so the agent can act on controls rather than guessing from pixels alone. Screenshots provide visual evidence, while request, response, console, and page-change data can support debugging.
Observation does not bypass permissions, login checks, verification steps, or site rules. Treat sensitive pages and destructive controls as approval boundaries.
For a catalog view, compare dsh-market. For evidence-gated task control, see J-Space.
Running Multiple Browser Sessions at Once
Browser instances can be named and a task can bind to a chosen profile. A session ID keeps subsequent commands attached to the same browser task.
- Start a session without stealing focus.
- Record the returned session ID and selected browser instance.
- Navigate and observe before interacting.
- Capture evidence after important state changes.
- Stop the session when the task is complete.
Example output contract:
Return the session ID, browser profile name, pages visited, actions taken, screenshots created, and any step that required human control.
Separate sessions reduce accidental cross-task state, but they do not create separate website accounts or permission scopes.
Installing BrowserSkill and the Live Web UI Overlay
Start with the official CLI installation guide. On macOS or Linux, install and verify bsk with:
curl -fsSLo /tmp/bsk-install.sh https://raw.githubusercontent.com/Tencent/BrowserSkill/3f10983c09fcd443c7420830ab984c9d409954bf/install.sh
EXPECTED_SHA256=d94f263caf9c2ef53240be3fc9988913fe4b82b25f056e4931448cad9d41f34d
if command -v sha256sum >/dev/null; then echo "$EXPECTED_SHA256 /tmp/bsk-install.sh" | sha256sum -c - || exit 1; else echo "$EXPECTED_SHA256 /tmp/bsk-install.sh" | shasum -a 256 -c - || exit 1; fi
sh /tmp/bsk-install.sh
rm /tmp/bsk-install.sh
export PATH="${BSK_INSTALL_DIR:-$HOME/.local/bin}:$PATH"
bsk --version
Install the Chrome extension or Microsoft Edge extension, enable the local connection, and then add the DSH plugin:
dsh plugin --profile web add @wxg-prc-cpg/browser-skill-dsh-plugin
dsh --profile web
Run bsk doctor, verify that the extension reports Connected, and confirm that DSH discovers the plugin with a harmless first task. A successful doctor result alone does not prove Skill discovery.
Browser-bridge status checked 2026-09-28: official DSH upstream still warns that its developer preview can introduce compatibility-breaking changes. Re-run bsk doctor after host updates.
Save Verified Browser Procedures With EvoMap
BrowserSkill runs browser procedures; EvoMap separately organizes verified agent assets. After approval boundaries, session details, observations, and results are checked, a team can explicitly package the procedure for reuse. See the EvoMap Wiki and compare last30days-skill-cn.
FAQ
Does BrowserSkill run headlessly?
Its main workflow uses a visible Chrome or Edge browser, and it can start an agent window without taking focus. Keep the visible window available for human takeover.
Does it reuse my browser login state?
It can work with existing accounts in the connected browser. Apply strict approval boundaries around sensitive actions.
Is the DSH plugin enough by itself?
No. The documented DSH setup also uses the bsk CLI and a connected browser extension. Compare dsh-market for plugin discovery inside DSH.
Can two sessions use different browser profiles?
The CLI supports named browser instances and selecting a browser when starting a session. Verify the intended profile before acting.
Preserve an Approved Browser Procedure
After a browser task is reproducible, preserve its approval boundaries, session details, observations, and verified result for reuse.