How Handback works

Handback closes the gap between noticing a bug and getting it fixed. You narrate the problem where it happens, the recording becomes a brief an agent can act on, and the fix comes back to you for a human sign-off. Five stages:

record

Narrate what you're seeing, in the app where it happens — browser recorder at /record, the Chrome extension, your phone's screen recorder via /phone, or drop an existing clip on /upload.

distill

Handback turns the recording into a repro-grade brief: a report written for agents, a timestamped transcript, deduplicated keyframes tiled into contact sheets, and the console/network errors that fired while you spoke.

pull

Your agent lists open walkthroughs over MCP and pulls the full brief — no copy-pasting context.

hand back

The agent posts its result — summary, PR link, files touched — onto the walkthrough's review thread and flips it to in review.

sign off

You watch the recording against the fix and approve, or send it back with a note. Nothing resolves without a human.

Setup takes about two minutes. First, sign up for Handback. Second, open /connect and mint your command — the page creates an API token and renders it into the exact line below. Third, paste that line into your terminal:

claude mcp add --transport http handback https://handback.dev/mcp --header "Authorization: Bearer hb_…"

The /connect page mints the token and renders this command with the token already in it — nothing to install, the server is hosted.

Then just ask: anything waiting on Handback? — the agent takes it from there.

Any MCP client works — Cursor, Windsurf, or your own agent. Claude Code is simply the path we test every day.

list_walkthroughs

Everything the token can reach, newest first, filterable by status.

get_walkthrough

The full brief: report, transcript, contact sheets, keyframes, video — presigned links, short-lived.

set_walkthrough_status

open → in_review → resolved as the work moves.

post_result

The agent's answer: summary, optional PR link and files touched. Lands on the review thread and notifies the uploader.

ask_reviewer

Hit a fork? The agent asks a clarifying question; the walkthrough waits in needs_info until a human answers.

attach_evidence

Before/after screenshots attached to the result, shown at sign-off.

Why the brief works: contact sheets — a model reading consecutive frames side by side follows what happened; the same frames one at a time, it doesn't.

You get a personal space by default. Create a team and invite reviewers from /team. Walkthroughs auto-file to projects by the origin they were recorded on. And you can share a polished cut with anyone via a link — watch pages need no account.

what leaves your machine

The recording, transcript and metadata you upload — nothing else. Full detail on the privacy page.

processors

Transcription runs on Groq, transcript cleanup on Anthropic, email on Resend. The extension also offers fully on-device transcription.

retention

Resolved walkthroughs auto-expire after 30 days (you can Keep one forever); raw takes from person-to-person edits purge 14 days after the final cut renders.

pricing

The free tier is live — 2 walkthroughs a month, each with cloud transcription and the refine pass. Billing on the paid plans hasn't opened yet; they're coming soon, priced per reviewer, not per walkthrough. See pricing.