The web that has no API
Internal tools, legacy portals, supplier and admin consoles. If you can reach it in a tab with your own login, it can be worked on — no integration to commission, no credentials to hand over.
Companion for Chrome
A pair of hands in your own browser
It reads the page, fills the form and works the tools that were never given an API — in your session, with your logins, and it stops at the step that should be yours to take.
No more copying the answer across
The work already lives in a tab — usually several. What has been missing is a way to act on them without narrating every step by hand, and a reason to trust that it will stop where you would have stopped.
Real clicks, not simulated ones
Input is delivered through the browser’s own debugging protocol, so React-controlled fields and forms that check for genuine events behave exactly as they do for you. Half-working automation is worse than none.
What the page actually said
Console output, network calls and the rendered DOM come back together, so a run can explain why a page did what it did rather than guessing from pixels.
Straight into the work
Findings land in the doc that asked for them, next to the rest of the task. Nothing is stranded in a side panel that empties when the browser restarts.
What people put it on
Supplier portals
Pull the week’s order status out of a portal with no export
Reads each order row, reconciles against the sheet, and flags the three that slipped.
Analytics
Lift the numbers a dashboard will not let you download
Reads the rendered figures, writes them into the metric model behind the task.
Admin consoles
Work a queue that only exists behind a login
Opens each item, applies the rule you described, and stops at anything it was not told to decide.
Research
Read twelve competitor pages into one comparison
Visits each in turn and returns a table, with every claim carrying the page it came from.
QA
Walk a flow and bring back the console
Drives the steps, captures errors and failed requests, and reports where it broke.
Records
Log the outcome where it belongs
Fills the fields you named in the system you already use, and shows you the diff first.
Stay in control
An agent with hands in your browser is only as good as the things it refuses to do. These are not settings buried behind a toggle; they are the default, and several of them cannot be switched off.
Three states per site. No master switch.
Every origin is allow, ask, or block — and nothing else. There is no global “skip all approvals” mode, because the one thing users report most often about such switches is that they break.
A blocklist that beats your own approval
Banking, payments, webmail, health and government sites are refused even if you approved them. Lifting that takes a separate, deliberate override on that one site.
It stops before anything consequential
A control that reads like a payment, a transfer or a deletion is refused outright. Releasing one such action takes a press in the extension popup, it lasts two minutes, and it is spent on the next attempt.
It never picks the tab
No tab groups are created, rearranged or adopted. A run acts on the tab it was given and detaches when it finishes.
Page text is evidence, never instructions
Everything read from a page is carried as untrusted data. A visited page cannot tell the run what to do next — the failure mode that makes browser agents dangerous.
A log, and a way out
Every action taken or refused is recorded and readable in the popup, which also shows what is attached and can detach everything in one press.
Worth saying plainly: treat any browser agent as you would a capable new colleague with your passwords. Keep it away from the workflows you would not hand over, read what it proposes, and stop it when something looks wrong.