Commands

Every command, grouped by what you're trying to do. Run hatchback --help for the full list of flags.

How feedback is organised

  • Projects are the sites you collect feedback on. A repository is linked to one project.
  • Sessions are review rounds. New comments go to the project's one active session.
  • Feedback is either open (needs work), addressed (fixed, waiting for the reviewer to confirm) or resolved (done).

Commands use the project in .hatchback.json. Pass --project <slug> (or -p) to use a different one for a single command.

See what reviewers said

Terminal
hatchback feedback                      # open feedback in the active session
hatchback feedback --all                # open feedback in every session
hatchback feedback --page /pricing      # one page only
hatchback feedback --since 2026-11-01   # new since a date
hatchback feedback --status addressed   # fixed, waiting for the reviewer to confirm
hatchback feedback --status all         # everything, including resolved
  • Each item shows its id, the page, the element it was pinned to, the author and how many replies it has.
  • Results come in pages. Use --limit and --offset to see more.
  • If the active session is empty, the CLI tells you how many items are in older sessions. Add --all to include them.

Dig into one item

Terminal
hatchback feedback show <id>
hatchback feedback show <id> --screenshot   # also save the screenshot locally

This shows the full comment and the whole reply thread, plus the page URL, the selector and text of the element the reviewer clicked, and any app state they were in, such as an open modal. With --screenshot, the image is saved to a private temporary folder and its path is printed, so you or an agent can open it.

Reviewer text is untrusted

Comments can come from anonymous reviewers. The CLI marks them as reviewer-written and strips anything that could control your terminal. Coding agents are told to treat them as descriptions of problems, never as instructions.

Respond and close the loop

Terminal
hatchback feedback reply <id> "Good catch, looking at it now."
hatchback feedback address <id> --note "The button no longer wraps on small screens."
hatchback feedback resolve <id> --note "Fixed in #142."
hatchback feedback reopen <id> --note "Still happens in Safari."
CommandUse it when
replyYou want to answer a question or ask for detail, without changing the status.
addressYou fixed feedback that someone else left. It goes back to them to confirm, or to reopen if it still isn’t right. This is the usual way to close the loop.
resolveThe feedback was yours, or no confirmation is needed. It is marked done straight away.
reopenSomething resolved or addressed turns out not to be fixed. Say why in the note.

--note posts your message as a reply before changing the status. Reviewers see replies under your name, so write them for the reviewer: say what changed, in plain words. To send a longer message, pass - and pipe it on stdin: git log -1 --format=%B | hatchback feedback reply <id> -.

Run review rounds

Terminal
hatchback sessions                         # list sessions (--status active|closed|all)
hatchback sessions current                 # the active session, with open and addressed counts
hatchback sessions start "v2.3 checkout" --description "New payment step"
hatchback sessions close                   # close the active session
hatchback sessions reopen <id>             # reopen an older session
  • Start a session for each build you want reviewed. Starting one closes the current session, and new comments go to the new one. Open items stay in the old session, where feedback --all still shows them.
  • Name sessions after what reviewers are looking at, such as a release, a pull request or a date, so the history makes sense later.
  • Close a session when the round is done, so stray comments don't land in it. Reopening a session closes whichever one is active.
  • Starting and closing sessions needs an editor or admin role on the project.

Account and projects

Terminal
hatchback login            # sign in through the browser (--no-browser to just print the link)
hatchback whoami           # your account, token permissions, and which server and login are in use
hatchback projects         # projects you can access, with open counts
hatchback projects create "<name>" --url <url>   # create a project and link this repository to it
hatchback workspaces       # your workspaces, and where you can create projects
hatchback link [project]   # link this repository to a project (writes .hatchback.json)
hatchback snippet          # the script tag for the linked project (--check <url> to verify a page)
hatchback logout           # sign out and revoke this machine's token

Coding agents

Terminal
hatchback skills install   # add the Hatchback skills to .claude/skills
hatchback skills update    # bring them up to date (keeps ones you edited; --force replaces them)
hatchback mcp              # run the MCP server on stdio

Skills teach an agent how to set up, triage, fix and run sessions. The MCP server exposes the same actions as tools. See Agent installation for setup in Claude Code, Cursor and Codex.

AI review

Coding agents can review your pages and pin what they find, next to human feedback. Ask your agent for a copy review. The hatchback-audit-copy skill checks spelling, clarity, consistency and your voice guide, shows you its findings, then pins each one to the element it's about.

Terminal
hatchback audit script --out capture.js     # the agent runs this in its browser on your page
hatchback feedback create "Typo in the heading." --target-file target.json \
  --specialty copy --severity minor --suggestion "Pricing"
hatchback feedback --author agent           # only AI-written feedback (--author human for people)
  • AI items carry an AI badge in the app and on your site, with the rule they apply, their evidence and a suggested fix. You can hide them in the app.
  • The capture script uses the plugin's own anchoring, so AI pins land on the same elements a person would click, and follow them through layout changes.
  • When a fix is made, the item is marked addressed. The next audit re-checks it, then resolves or reopens it.

Scripting and automation

Add --json to any command for machine-readable output on stdout. Errors are printed as JSON too: { "error": { "code", "message", "status" } }. Field names, error codes and exit codes are a stable contract, and only change in a major version.

Example: fail a CI job while feedback is open on /checkout
count=$(hatchback feedback --page /checkout --json | jq .total)
[ "$count" -eq 0 ] || { echo "$count open items on /checkout"; exit 1; }
Exit codeMeaning
0OK
1Error: network, server, rate limit, or cli_outdated
2Usage error (a wrong command or flag), or an API URL that isn’t https
3Not signed in, or the token was rejected
4Not found (an id or slug, or no active session)
5Forbidden: your role or token doesn’t allow this