box-desktop
When a task needs your own desktop or browser, such as a website with no connector, a GUI app, or a sign-in only the user can complete, read before you dispatch the first browser or desktop subagent.
SKILL.mdnon-stock/home/box/agent-data/managed-skills/skills/box-desktop/SKILL.md
GitHubThe box desktop
You hold the read-only Screenshot tool to see its current screen, confirm where a flow landed, or check on a running subagent. You cannot click, move, type, press keys, scroll, or wait on the desktop yourself. Delegate every browser and desktop interaction to a subagent. Do not bypass this boundary with Shell-driven GUI automation such as xdotool, or by driving the box browser from Shell: no CDP attach, no Playwright, Puppeteer, or websocket-client, no /json/new, no cookie-DB scraping, and no page JS eval over DevTools.
Delegate the outcome, constraints, required values, and success criteria, not browser or desktop steps. Prescribe a modality only when that modality is itself part of the desired result; otherwise let the child choose.
When you know the destination URL, whether one the user pasted or one you can construct (a site's search/filter URL like
https://www.amazon.com/s?k=bread+flour), put that exact URL in the task, as specific as the site's query params allow, so the subagent opens it directly instead of clicking through the site to rebuild it.For bulk or structured data, don't type it in by hand: generate the file with Shell (e.g. a CSV), inspect it with Read when useful, then have the subagent import or upload it, far faster and more reliable than entering values one by one.
When it returns, read its report before acting. If it stopped short or hit a step only the user can do, that's your cue to follow up or hand off the box.
When
request_cookie_origin_approvalis among your tools, try the user's Chrome cookies first for a sign-in. List origins, and if the site is listed, request it. Only when that path is exhausted (not listed, denied, or page still wants a fresh login) do the sign-in in the box browser. Follow that tool's approval and site-exception rules.When a page needs the USER to type (login, address, phone, OTP) and
request_user_formis among your tools, Read and follow thein-chat-formsskill first. Have the child open the page and report fresh snapshot targets; use a form only when fields are fillable. Do not jump torequest_box_helpjust because a site needs a password.For steps that need the user, use
request_box_helpwhenrequest_user_formis not offered or a form cannot express or reach the step, including non-text/native steps. A structural preflight refusal is final, so don't re-issue the same form. Don't pre-ask "hand you the box?" The handoff is the ask.