Give your coding agent real client feedback.
Retyping a client email into a prompt strips out the context and adds your guesswork. Let the agent read the client's own pins instead: same words, same element, same screenshot.
The workflow
- Client pins the live page. Invite them as a free commenter, or embed the widget on a staging URL. They click and type; there is nothing to learn.
- You approve. Not every client comment should become a task.
approve_pinmarks the ones the agent should build. - Agent reads the packets. In Cursor, Claude Code or Codex with the MCP server connected,
implement_approved_pinsreturns approved pins grouped by page, with selectors, threads and a suggested branch. - Agent claims and fixes.
claim_pinbefore the change,fix_and_resolveafter, with commit SHA, branch and PR URL. - Client sees the result. The resolved pin carries the trail, so you are not writing "done, please check" emails.
Keep a human in the loop
Review the diff as you would any agent change. When a pin is ambiguous, the agent can ask on the thread with add_agent_reply and the client answers in place. Check the fix on the real page before you tell the client it is live.
Go deeper
Editor-specific setup is in Pincushion for Cursor and Pincushion for Claude Code. The full walkthrough is client feedback into Cursor, and the tool-level view is handing visual feedback to AI coding agents.
Start free
Pincushion's free tier covers one project with unlimited manual pins and 25 AI actions a month, and every new account starts with a 14-day full-Pro trial. Commenters are always free. See pricing for the details (checked 2026-10-28).