Browser control is exposed to the agent as mcp__chrome__* tools. Once the extension connects there are 27 of them, grouped by capability rather than being one unrestricted interface.
The tools are published by the DSH plugin dsh-chrome-control and executed by the Chrome extension; they only appear once both are installed — see Install the Chrome extension.
Sessions and tabs
| Tool | What it does |
|---|
find_tab | re-select a tab by URL, or borrow the one the user is viewing |
list_tabs | list every tab in this session’s group with id, URL and title |
close_tab | close the current tab |
close_session | close the whole tab group (only when explicitly asked) |
Navigation and reading
| Tool | What it does |
|---|
navigate | open a URL |
get_text | extract readable text, stripping nav, header, footer and scripts |
snapshot | read the accessibility tree for interactive elements and @e refs |
Targeting and interaction
| Tool | What it does |
|---|
find | locate a control from a plain-language query, returning reusable refs |
click | click an element; falls back to real input when a widget ignores synthetic events |
fill | set an input, textarea or rich editor (clear, insert, then read back) |
select | choose a dropdown option: native <select> or an ARIA combobox |
hover | real pointer hover, firing hover-only menus and tooltips |
focus | focus without clicking, so you can type next |
key_type | type text with real key events and verify by reading back |
send_keys | press one key or chord (such as Control+A, Enter) |
mouse_click | trusted mouse click by element or viewport coordinate |
upload | set local file paths on an upload control |
| Tool | What it does |
|---|
scroll | scroll the page or one scrollable element, reporting whether it moved |
scroll_into_view | bring an element into view and return its geometry |
Scripting and evidence
| Tool | What it does |
|---|
evaluate | run JavaScript in the page and return a serializable result |
screenshot | capture the viewport, or one element |
save_as_pdf | render the page to a PDF file on disk |
network | list recent requests for this tab (method, URL, status) |
network_detail | full record of one request, with headers and optional body |
dialog | answer alert / confirm / beforeunload / prompt |
Waiting
| Tool | What it does |
|---|
wait | sleep a fixed interval (capped at three seconds) |
wait_for_selector | poll until an element is visible, or gone |
It reuses your session
These tools run inside the Chrome you are already signed in to, so reaching authenticated sites needs no extra login. The trade-off is worth stating plainly: the agent can see what you can see in that browser.
Next
- Permissions — default-deny and the allowlist for plugin routes.
