Co Browser

co browser Skill

Drive one real browser from the shell. The browser stays open between commands — cookies and logins persist until co browser close. Every terminal on the machine talks to the same daemon: one browser, one board, no matter where you type.

Always read the output, not just the exit code. A failed action (selector not found, browser not open) often returns exit 0 with the error text on stdout — never chain co browser ... && next_step as your only success check.

Step 1: Identity — who are you?

The daemon attributes every tab to a caller and uses that identity to stop two agents from silently driving the same page. Set CO_WHO explicitly — shell state does not survive between your tool calls, so a lone export is lost. Either prefix every command:

CO_WHO=alice co browser go_to example.com

or keep the export and the commands in the SAME shell invocation:

export CO_WHO=alice && co browser go_to example.com && co browser get_text

(Some environments — e.g. Claude Code background jobs — are auto-identified, but setting CO_WHO is always safe and never worse.) An anonymous caller still works but gets no contention protection.

Step 2: Check the board, then claim a tab

One task = one tab. Before touching the browser, see what's already happening:

co browser status      # browser state, headless flag, last command, the board
co browser tab ls      # every tab: who owns it, purpose, last command (--json for scripting)

Solo (nobody else on the board) — bare commands run on the shared main tab, no ceremony:

co browser go_to example.com     # runs on 'main'
co browser get_text              # still 'main'

Concurrent (another agent or a second parallel task exists) — open your own tab and add -t <name> to EVERY command, including do. Pick an explicit literal tab name — a NAME=$(...) capture is lost between tool calls:

CO_WHO=alice co browser tab open scrape --for "scrape pricing" --needs 10m
CO_WHO=alice co browser -t scrape go_to example.com/pricing
CO_WHO=alice co browser -t scrape do "extract every plan and its monthly price"
CO_WHO=alice co browser tab close scrape       # release when done

--needs is your estimate of how long you will hold the tab (30s, 10m, 2h). Say it: other agents leave your tab alone until then, and co browser tab ls shows them when you expect to finish. Without it, two minutes of silence is enough for another agent to take your tab — which is wrong when you are waiting on a slow page or a human.

When delegating browser work to subagents, write each subagent's tab name into its prompt — parallel agents share the one logged-in browser through separate tabs.

When the site opens its own tab — a target="_blank" link, a "view invoice" button, a payment popup — that page belongs to no session, so tab ls cannot show it as yours and every -t command keeps running in the page you were on. tab ls tells you one exists; these two reach it:

co browser list_pages                  # every real page, with who is driving it
co browser -t mytask switch_page 1     # your session now drives page 1
co browser -t mytask get_text          # ...and every command after it
co browser -t mytask switch_page 0     # back to where you were

Indexes come from list_pages and shift when pages open or close, so read the list again rather than reusing a number from earlier in the task. A page another session is driving is refused by name — two agents cannot share one page, the same rule as tabs.

Step 3: Pick the right verb

Two ways to drive the browser — pick per command, mix freely:

Direct functions (deterministic, instant, free — no LLM). Use for every step you can spell out:

co browser go_to https://example.com/login        # https:// assumed if omitted
co browser get_text                               # visible page text
co browser get_links_from_page                    # every link, one per line
co browser take_screenshot /tmp/page.png          # saves a PNG, prints its path
co browser click_element_by_selector "button" --index=0 --text="Send message"
co browser type_text_by_selector "#email" "user@example.com"
co browser fill_text_by_selector "#invite" --stdin < invite.txt
co browser scroll                                 # scroll the page (scroll 3 = fewer steps)
co browser get_current_url

co browser help prints the live list of every function with its arguments — trust it over any example here. Calling a function with wrong arguments returns its usage line: self-correct from that.

Typing into whatever already has focus — after a mouse_click into a field that no selector reaches cleanly — is keyboard_type, and pressing keys is keyboard_press. There is no type_text; that name returns "unknown command" and costs a round trip. Use fill_text_by_selector to replace a controlled input atomically, type_text_by_selector for humanized appending when a selector works, and keyboard_type when only focus is available.

keyboard_type appends — it does not replace. To overwrite a field: click into it, then keyboard_press "Meta+a" (macOS) and keyboard_press "Backspace" before typing, or you get 1:15 AM09:00 AM. And confirm the click actually landed inside the input: if it missed, Meta+a selects the entire page instead and your text goes nowhere while the screen turns blue. The screenshot after typing is the only way to tell those two outcomes apart. (The verb is scroll, not scroll_down; guessing a plausible name costs a round trip.)

Never hunt for an item by scrolling a feed. Infinite lists re-render as they lazy-load, so scroll position is not reproducible: the same scroll 6 lands somewhere different each time, Home/Meta+Home do not return you to the top, and the card you were aiming at moves between screenshots. Worse, clicking a per-item menu you located by eye can hit a neighbouring item — on a destructive action that is the wrong row deleted.

Extract the item's id once, then address it directly:

# 1. one script over the whole list, returning ids -- not screenshots
co browser -t mytab run_page_script /abs/find-id.js     # match on text, return data-urn/data-id
# 2. open that item alone, where its controls cannot be confused with anything else
co browser -t mytab go_to "https://example.com/item/<id>/"

Six scroll-and-screenshot round trips became one script plus one go_to this way, and the per-item menu became unambiguous because the page held exactly one item.

do "<instruction>" (natural language — an AI agent sees the page and works out the steps). Use for judgment, not for steps you already know:

co browser do "log in with the saved credentials and open my notifications"

Describe the end state, not the steps. A do is a full agent run — many LLM calls, possibly minutes — and the daemon is busy for its whole duration, so other commands queue behind it. Never wrap do in timeout: killing the client does not stop the run, it only orphans it (you pay for an answer nobody reads).

For state checks, prefer the free text probes first (exit 0 = matched):

co browser get_current_url | grep "/feed"

Only when the state is visually ambiguous, fall back to a one-word do probe:

co browser do "Look at the current page. Reply EXACTLY one of: STATE: feed | STATE: login | STATE: error"

Step 4: Verify with evidence

• Routine checks are text checks: get_current_url / get_text after navigation and clicks. Reserve screenshots for evidence gates.

• Irreversible actions (submit, post, publish, pay) get the full gate: screenshot before, act exactly once, screenshot after. Never click the same submit button twice — if the result is ambiguous, report uncertain state instead of retrying.

• Long pages: dump once and read locally instead of re-extracting per chunk:

  co browser get_text > /tmp/page.txt    # one daemon round-trip
  sed -n '1,200p' /tmp/page.txt
  

• To find stable selectors, save the DOM and grep it:

  co browser save_page_context li_composer
  grep -o 'aria-label="[^"]*"' ~/.co/browser_context/*li_composer*/page.html | sort -u
  

• Site-specific DOM logic belongs in skill-local JS run with an absolute path (the daemon resolves relative paths against its own cwd, not yours):

  co browser run_page_script /path/to/project/.co/skills/<skill>/scripts/extract-items.js
  

Step 5: React to failures

Exit codes cover the daemon-level contract; the output text covers the action itself:

The exit-3/4 error messages ARE the documentation — they always carry the current recovery steps, so trust them over this table.

When a tab blocks you, read co browser tab ls before deciding. Each tab shows when its owner expects to finish:

[scrape] https://...  who=alice  purpose='scrape pricing'  open 3m
   owner expects to finish by 14:20 (7m left) — leave it alone until then

Inside that window, open your own tab instead. Once it has passed, the tab is free and closing it is a courtesy — an estimate that ran out with the tab still open means that agent crashed, not that it is still working. Every agent here is cooperative; tidying up after a dead peer is expected, taking a live agent's page is not.

Logging in

The default is a visible window — exactly what you want for a first login. Never automate credentials: open the login page, tell the user, then check with short waits — one short check per tool call, not one long blocking loop (a long loop hits the tool timeout, and every poll refreshes your claim on the tab, locking other agents out for its whole duration):

co browser go_to https://site.com/login
# tell the user to log in in the visible window, then per check-in:
sleep 15 && co browser get_current_url | tail -1

Repeat the check a handful of times; if the URL is still the login page after that, stop and ask the user instead of looping forever. Logins live in the profile (~/.co/browser_profile/), not the daemon — they survive close and restarts. Log in once, stay logged in.

Headless: co browser --headless go_to ... — the mode is fixed by whichever command starts the daemon (status shows headless=true/false). To switch: co browser close, then relaunch.

Scripting hygiene

• Wrap direct functions that might block in timeout 60 ...; never timeout a do.
• Take the data line with tail -1 when output includes env banners.
• Batch related commands in one tool call; never spend a whole call on a bare wait.
• wait is in seconds — every other settle knob here is milliseconds (--wait_ms=2500), so wait 2500 looks right and means 41 minutes. It is capped at 60s and refuses anything longer. For a settle longer than a second or two you want a condition, not a sleep: wait_for_element / wait_for_text.
• Uploads: try upload_file_by_selector 'input[type="file"]' <path> first; if the input is hidden behind a button, upload_file_after_click_by_selector '<button selector>' <path>.

Troubleshooting

• "Where is my window?" — co browser status says headless=true? An earlier command started the daemon headless. co browser close, rerun without the flag.

• "daemon is busy" — a long do is holding the single-threaded daemon. Wait and retry; co browser status shows the last command once it frees up.

• "Opening in existing browser session" / profile in use (shows a PID) — a manually opened Chrome or stale daemon holds the profile; kill that PID, retry.

• TargetClosedError after a crash — the page died under the daemon; run co browser open_browser again before retrying the command.

• "Chrome failed to start" — usually ssh/cron without a desktop session (run from a logged-in Terminal, or use --headless). Full launch log: ~/.co/browser.log.

• Nuclear option — kill the daemon and let the next command start fresh (logins survive in the profile):

  pkill -f 'connectonion.cli.browser_agent[.]daemon'
  

Done checklist

• Identity attached to every command (CO_WHO=... co browser ...)
• Board checked (status / tab ls) before claiming a tab
• Own tab used if any other agent/task shares the browser (-t on every command)
• Output text read on every command, not just exit codes
• Irreversible actions performed exactly once, with before/after screenshots
• Tab closed (co browser tab close <name>) when the task is done
• Browser left open for the next task unless the user asked to close