Chrome Relay

Drives the user's real Chrome through a Chrome extension + local native host. Prefer it when logged-in browser state (auth cookies, sessions, installed extensions) matters.

Setup

1. Chrome extension
2. CLI:

``sh npm install -g chrome-relay@latest chrome-relay install chrome-relay --version chrome-relay doctor ``

Use the latest CLI. Multi-browser/profile routing and uploads require >= 0.8.0; Dia detection and the agent-friendly profile picker require >= 0.8.1. If the printed version stays below 0.8 after installing, stop and resolve the stale binary on PATH before using 0.8 commands:

chrome-relay --version
which -a chrome-relay        # macOS/Linux; use `where chrome-relay` on Windows

At the start of a session, run chrome-relay profile list. With one connected instance, normal commands need no profile flag. With several, this tells you exactly which browsers and profiles are reachable before you act.

The core loop

chrome-relay tabs                             # find or create a tab
chrome-relay navigate "https://kushalsm.com" --new   # background tab by default
chrome-relay snapshot --tab 1234 -i           # see the page: actionable elements get @refs
chrome-relay click @e12                       # act on refs, no --tab, no selector
chrome-relay fill @e14 "hello"
chrome-relay wait --text "Saved" --tab 1234   # block until the page reacts
chrome-relay snapshot --tab 1234 --diff       # print only what changed (~100 tokens)

Snapshot output is compact indented text, usually 1 to 15 KB for most pages. Read it directly, no jq needed:

- link "Hacker News" [ref=e4]
- textbox "Search" [ref=e41]: current value
- checkbox "Remember me" [checked, ref=e42]
- clickable "Open card" [ref=e88]        # cursor-pointer div the AX tree missed

Refs carry their own tab. click @e12 acts on the tab that produced e12, never the active tab. Safe while the user keeps browsing. A contradicting --tab errors with target_conflict.

Ref lifetime. Refs survive same-page DOM churn (cached backendNodeId, healed by role+name re-find when nodes are replaced) but die on real navigation. A dead ref returns error.code = stale_ref. Re-run snapshot.

Interception. Ref clicks hit-test the point first. If an overlay / sticky header / modal owns it, you get error.code = click_intercepted naming the interceptor. Dismiss it or scroll, then retry. The click was NOT delivered. fill/type skip this check (covered inputs are still writable).

Tool surface

Picking the right text tool

Many browsers & profiles (CLI >= 0.8)

The primary supported targets are Google Chrome (including multiple Chrome profiles), Dia, and Brave. One CLI reaches every connected instance where the extension is installed; each Chrome profile is a separate addressable instance.

The installer also knows native-host manifest paths for Chrome Canary, Chromium, Edge, Vivaldi, Arc, and Opera. Treat those as compatibility targets unless the current task has verified them; do not claim that manifest detection alone proves full browser support.

Install the extension once in every browser/profile you want reachable, then run chrome-relay install once so every detected browser can spawn its own host.

chrome-relay profile list             # who's connected: label, browser, id prefix
chrome-relay profile label work       # one connected: label it directly
chrome-relay --profile 3f2a profile label personal  # several: first pick by id prefix
chrome-relay --profile work tabs      # scope any command (global or per-command flag)
chrome-relay click @3f2a:e12          # snapshot refs are profile-qualified and route by THEMSELVES

One instance connected: no flags, everything routes implicitly. Several: unscoped commands fail profile_ambiguous. Treat the error as a picker: choose one of its exact --profile <label|idprefix> entries and rerun the command. It never guesses. Refs carry their profile the way they carry their tab, so after one snapshot you rarely need the flag again. Free a stale label with chrome-relay profile unlabel <name>.

Uploads (CLI >= 0.8)

Three strategies, no auto-fallback — the failure names the strategy, pick the next:

chrome-relay upload set --selector 'input[type=file]' --tab 123 ./cv.pdf   # direct; works on HIDDEN inputs
chrome-relay upload choose --click-ref @e4 ./cv.pdf   # trigger opens the OS picker: intercepted, NO dialog appears
chrome-relay upload drop --selector '.dropzone' ./avatar.png

Files are paths — Chrome reads them itself, no size caps. notafileinput → use choose. nofilechooser → wrong trigger, or it's a drop zone. fileaccess_denied → chrome://extensions → Chrome Relay → "Allow access to file URLs". set/choose return what the input ACTUALLY holds after the call.

Element addressing: the fallback ladder

1. @ref from snapshot -i: default. Covers buttons/links/inputs, named content, cursor-pointer div-soup (the sweep), and shadow DOM.
2. CSS selector: when you know the selector statically and don't need a snapshot.
3. js probe, then coordinate click: canvas internals and SVG chart segments (anonymous <path> elements have no DOM handle anywhere):

``sh chrome-relay js --tab 1234 "const r = document.querySelector('svg path').getBoundingClientRect(); return {x: r.x + r.width/2, y: r.y + r.height/2}" chrome-relay click --tab 1234 --x 312 --y 218 ``

Don't poll. Wait.

A snapshot after every action wastes turns. The cheap loop on a changing page:

chrome-relay click @e12
chrome-relay wait --text "Saved" --tab 1234     # or wait <selector> / --url / --load
chrome-relay snapshot --tab 1234 --diff         # only the changes, refs included

Top gotchas

1. snapshot -i is for ACTING, not fact extraction. It prints ref-bearing elements only. Non-interactive values (dashboard metrics, paragraph text, chart labels) drop out. Measured live: a Cloudflare Pages metrics page lost all its numbers under -i. To READ facts, use full snapshot, get text <target>, or a js projection.
2. type appends. It inserts at the caret. If the input had a value (autosaved draft, default text), clear it first via js or keys (Cmd+A then Backspace).
3. Refs die on navigation. stale_ref means the page changed under you; re-snapshot. Don't retry the same ref.
4. Coords go stale fast. Read getBoundingClientRect, scroll/reflow, then click, and you hit the wrong element. For autocomplete popups especially, use keyboard nav, not coord clicks.
5. Click "succeeded" but nothing happened. First diagnostic: document.elementFromPoint(x, y). If it returns a wrapper or form background, your coords are wrong. If it returns the right element but state didn't change, you're likely on chrome-relay <0.5.20. Upgrade.

More recipes: [references/patterns.md](references/patterns.md) Failure modes: [references/troubleshooting.md](references/troubleshooting.md)

Operational guidance

• Don't give up early. A failing click is information, not a stop signal. Attach a document-level listener with capture:true and watch what fires:

``sh chrome-relay js --tab 1234 " ['pointerdown','mousedown','click'].forEach(t => document.addEventListener(t, e => console.log(t, e.target.tagName, e.target.className), {capture:true}) ); return 'listening' " # do the action, then: chrome-relay console --tab 1234 ``

• Don't echo secrets. When extracting tokens / API keys via js, write the result directly to a file. Never echo $TOKEN or interpolate into shell strings. It ends up in scrollback, logs, and tool transcripts.
• Redact network output. Request/response headers carry cookies, auth/CSRF tokens, account and project IDs. Never paste raw chrome-relay network output into chat, docs, issues, or commits. Filter to the fields you need (url, status, timings) or redact headers first.
• Capture before irreversible actions (form submit, send message, account change). Save the screenshot path.

Guardrails

• Errors are structured: branch on relayError.code (staleref, clickintercepted, elementnotfound, targetconflict, profileambiguous, timeout), not on message text.
• If a flag is unclear, chrome-relay <command> --help is authoritative. These docs lag.