Let AI agents drive your real Chrome browser — logged-in tabs, no relaunch, zero dependencies
Chrome-bridge exposes 26 tools with inconsistent quality. Most tools have descriptions (good), but schemas are incomplete or missing type information. Parameter descriptions exist but lack constraints (enums, ranges, formats). No output schemas documented. Error handling is minimal, responses don't guide recovery. Tool naming is verb-based (snap, click, fill) but some names are domain-specific jargon ('snap' for accessibility tree, '@eN refs') that won't be self-evident to LLMs unfamiliar with the bridge's conventions. Composition is reasonable, tools chain well (snap → click/fill), but many tools accept a 'match' parameter that is overloaded (URL substring, title substring, or id:<tabId>), forcing LLMs to reason about multiple input formats. No batch variants despite agents likely calling tools in loops.
read commands from stdin, one per line ('#' = comment, quotes honored; eval payloads ride verbatim — quotes are JS syntax there, not grouping) — one process for N commands; stops on first error
click an element (snap first to find @eN refs, or use CSS)
close the matching tab — --all closes every match (the remedy for identical-URL tabs no <match> can separate)
read the page's console output (logs, errors, warnings)
browser-hygiene sweep: tabs wearing bridge markers the bridge doesn't track (🟣 group, status favicon, emulation, debugger), each with its last lifecycle events; --fix releases every residue tab and prunes dead-id memory (driven tabs are never touched)
emulate a device (mobile, tablet, desktop with custom viewport)
Overloaded 'match' parameter across 15+ tools accepts multiple formats (URL substring, title substring, id:<tabId>) without enum constraints or clear format documentation. LLMs cannot determine which format to use without trial-and-error.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., tab_id from tabs response). Responses are described in prose but not formally specified.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
evaluate JavaScript in the page context (returns the result)
fetch a URL from the page context (respects cookies, auth, CORS)
fill a form field (snap first to find @eN refs, or use CSS)
find text on the page (returns @eN refs of matching elements)
check bridge server and extension health status
read the command history log
measure element dimensions and position
navigate matching tab (waits for load, 8s cap; same loaded:false semantics as open)
capture network requests (HAR format with --har)
open + mark a new tab (waits for load, 8s cap; the reply's loaded:false means the cap fired on a still-loading page — snap/eval/wait --text work on what's there)
paste text into the focused element (reads from clipboard or --text argument)
list connected Chrome profiles — id and name (for --profile) + version
release a driven tab (stop driving it)
take a screenshot (PNG data URL or file path with --out)
a11y-tree snapshot with @eN refs (cheap — use before shot); [css|@ref] scopes to a subtree (@ref = the --skeleton drill-down), --diff shows only changes since last snap, --href includes all link URLs (default: only nameless links); --skeleton: depth-limited map — cut containers read '… N inside'
start the bridge server
stop the bridge server
read the service worker's console log ring
list tabs (compact JSON); [match] filters by URL/title substring; with multiple Chrome profiles connected, merged with a profile tag. Rows carry the tab id: a <match> of id:<tabId> — from the toolbar popup's Copy button — targets exactly that tab, no ambiguity (ids die on browser restart and change on prerender — re-copy)
wait for a condition (text, selector, or time)
Domain-specific jargon in descriptions ('snap' for a11y-tree, '@eN refs', 'skeleton drill-down', 'bridge markers') assumes LLM familiarity with chrome-bridge internals. Descriptions should explain what the tool does in plain language before introducing domain terms.
No error handling guidance. Tool descriptions do not explain what errors are possible, when to retry, or how to recover. E.g., 'open' says 'loaded:false means the cap fired' but does not advise the agent what to do next.
Parameter 'profile' is optional across all tools but descriptions do not explain the default behavior when omitted (e.g., does it auto-select if one profile is connected? Error if multiple?). Undocumented dependencies cause silent misuse.