| name | browser-control-basics |
| description | Use Chromium safely and predictably in this environment. Covers MCP versus direct CDP, what counts as real browser control, and the minimum verification/reporting rules. |
| license | CC0-1.0 |
Browser Control Basics
Use this skill when a task involves driving, inspecting, or validating behavior
in Chromium.
This skill defines the shared environment rules for browser control. Load it
before using more specialized browser skills.
Expected browser
This setup expects browser automation to use the Chromium instance launched by:
chromium-agent
Do not launch raw chromium, google-chrome, or a fresh browser profile for
agent-controlled browser work unless explicitly instructed.
Prefer the existing running Chromium session over launching a separate browser
stack.
Environment note
In this environment, chrome-devtools MCP may be unavailable or disconnected
even when Chromium is running and reachable over CDP.
Treat MCP availability and browser controllability as separate concerns.
Startup check
Before debugging MCP itself, first check whether the expected browser is running
and reachable.
pgrep -af 'chromium.*remote-debugging|chromium-agent'
Check the CDP endpoint used by this setup:
curl -s http://127.0.0.1:9222/json/version
Adjust the port only if the local wrapper uses a different one.
If the expected browser is not running, launch it with:
chromium-agent
Allowed control paths
If the user asks for browser automation or inspection:
- use MCP when it is available and appropriate
- if MCP is unavailable, direct CDP fallback is allowed when possible
- if both MCP and CDP fail, check
chromium-agent before debugging the MCP
server
What does not prove control
Do not assume browser control just because:
- a Chromium process exists
- a URL opens in the browser
- a page is visible on screen
Real browser control should be verified with an actual inspection or browser
action, such as:
- listing or selecting targets
- reading DOM state, console output, or network failures
- evaluating script in the page
- clicking, typing, navigating, or taking a screenshot
Reporting rule
Briefly state which control path you used:
- MCP
- direct CDP fallback
- no browser control available
Scope discipline
Use the smallest capable control path for the task.
Do not build a fresh automation stack unless:
- browser control paths are unavailable
- the task genuinely needs a dedicated scripted runner
- repeated execution makes scripting clearly worthwhile