It reads the room
get_project_context returns the plan and the human's live view — the open sprint, the active filter. The agent proposes into a context, not a vacuum.
Built on WebMCP
Your browser has WebMCP switched off, so the tools on this page are dormant. Everything below still works — turn the flag on and reload to watch them arrive.
What would you like to do?
Measure before rebuilding
The brief says drop-off is around 40%, but nothing records it per step. I put a short instrumentation sprint in front and re-sized the route collapse.
sprintSprint 0 — Instrument first
Know the drop-off rate per step before changing anything.
taskAdd funnel events for each checkout step
Sprint 0 — Instrument first (new) · 4h
requirementCheckout drop-off is measurable per step
high priority
taskCollapse the three checkout steps into one route
get_project_context returns the plan and the human's live view — the open sprint, the active filter. The agent proposes into a context, not a vacuum.
Every write tool authors a change set. Additions, field-level changes, removals. A sprint and the twelve tasks inside it arrive as one reviewable unit.
apply_pending_changes blocks on your decision. Approve and it lands in one transaction. Reject and your reason goes back as the tool result.
The set changes with the page. The projects list offers list_projects and create_project instead — which is what toolchange is for.
It needs desktop Chrome. The badge in the header turns on and lists the live tools.
To drive it with ChatGPT, use its Chrome extension in that same browser — or skip the flag entirely and open this page in the ChatGPT desktop app, in ChatGPT mode: its built-in browser exposes WebMCP on its own. Codex mode in that same app does not, so if the tools are dormant there, that is why.