Use case · anyone shipping with an agent

Visual feedback tool for AI coding agents.

Point at what's wrong on the page in front of you. Handoff captures the element, the screenshot, the selector, the computed styles and the console — and hands the whole thing to your agent over MCP. No package in your project, no access to the source.

Half your prompt is
describing what you're looking at.

writing three paragraphs about a broken layout
“the button under the second card, on the right”
pasting a screenshot and hoping the agent finds the element
re-explaining the same page in every new chat

Just point at it.

Install once.
Works on any page.

1 · It's an extension, not a dependency

Nothing to add to your project, so it works where a package can't: on production, on staging, behind a login, on a client's site whose code you've never seen.

2 · Click the element, leave a note

The pin sticks to the element, not to a coordinate on a screenshot. Add a severity and a sentence — that's the whole ritual.

3 · Your agent reads it over MCP

One HTTP endpoint, standard Bearer auth. Claude Code, Cursor, Codex or any other MCP client picks up every open pin without an export step.

4 · The fix comes back on the same wire

The agent replies on the pin's thread with what it changed and marks it resolved. You accept it or reopen it — the loop closes where it started.

One pin carries more
than a paragraph of prompt.

Everything a developer would open DevTools for travels with the note automatically. The agent doesn't ask “which button?” because it already has the selector.

Claude Code · via MCP: “Raised the button contrast to 4.6:1 (AA pass) in Button.tsx. Commit a3f9c2.”

Real pins, on a real page.

Four findings on a demo product we run, opened here through the same read-only link a teammate would get. Each one carries the element's screenshot, its selector and the styles that were computed at the time.

Filed by an agent over MCP. Nothing here is staged: the contrast really is 1.66:1.

What is visual feedback for AI agents?

Visual feedback for AI agents is a bug report written by pointing instead of typing. Instead of describing an element in prose, you click it on the running page; the tool captures what a developer would have opened DevTools to find, and delivers it to the agent in a form it can act on.

The distinction that matters is the instance. An agent can reason about your code, but it cannot see the page your user is looking at: which element, in which state, at which width, with which console error underneath. That context lives in the browser, not in the repository, and until something carries it across, the person has to retype it every time.

Handoff carries it as a pin: element, screenshot, selector, computed styles, HTML, console output, URL and viewport, delivered over MCP. Because it's a browser extension rather than a package, the page doesn't have to be yours — production, staging and sites behind a login all work the same way.

Questions

Which AI agents work?

Any MCP client: Claude Code, Cursor, Codex and others. One HTTP endpoint, standard Bearer auth — no plugin per agent.

Does it work on a site I don't own?

Yes. Handoff runs as a browser extension, so it needs the page, not the source. Production, staging, a client's site, a page behind a login — all the same.

Do my teammates need the extension?

Only to leave pins. Following a review works through a shared link in any browser, with no account and nothing installed.

Does Handoff read my codebase?

No. The MCP server serves your feedback — pins, context, screenshots. Your agent does the fixing locally, in your repository.

Is it really free?

Yes — unlimited reviews and unlimited seats during the public beta. No credit card.

Stop describing. Start pointing.