Agent installation
Let your coding agent install and set up the CLI. Then ask it to triage the feedback, fix it and reply to the reviewers.
Set it up with one prompt
Paste this into Claude Code, Cursor, Codex or any agent that can run commands, in the repository for your site:
Set up the Hatchback CLI for this repository. Fetch https://hatchback.app/agents/cli.md and follow the instructions step by step. Ask me before anything that needs my decision or my browser.The prompt points the agent at step-by-step instructions written for agents. The agent will:
- Check Node.js and install the CLI, asking first whether to install it globally or use
npx. - Check whether you're signed in. If not, it asks you to run
hatchback login. Agents never sign in on your behalf. - Link the repository to the right project, asking you which one if it isn't obvious. If there isn't one, it can create it and add the script tag to your site.
- Install the skills, add an
AGENTS.mdsnippet, and set up the MCP server if you want it. - Check that everything works and tell you how much open feedback there is.
Your part
hatchback login in a terminal and approve the sign-in in your browser. In Claude Code you can do it without leaving the session: type ! followed by the command.Then ask for work
- “Create a Hatchback project for this prototype and add the snippet.”
- “What feedback came in on the new pricing page?”
- “Triage the Hatchback feedback and propose a plan.”
- “Fix the layout issues reviewers found and hand them back for confirmation.”
- “Start a new review session for PR #42.”
After a fix, agents mark the item addressed with a note for the reviewer, who confirms it or reopens it. They only resolve feedback outright when you wrote it, and they ask before acting on many items at once.
Manual setup by agent
Prefer to set it up yourself? Install the CLI, sign in and link the repository (see Installation), then connect your agent.
Claude Code
Install the skills into the repository. Claude Code loads them when they're relevant:
hatchback skills installOptionally, add the MCP server too:
claude mcp add hatchback -- npx -y hatchback-tools@latest mcpCursor
{
"mcpServers": {
"hatchback": { "command": "npx", "args": ["-y", "hatchback-tools@latest", "mcp"] }
}
}Codex
[mcp_servers.hatchback]
command = "npx"
args = ["-y", "hatchback-tools@latest", "mcp"]Any agent that reads AGENTS.md
hatchback skills install prints a short snippet. Paste it into the repository's AGENTS.md so the agent knows the CLI is there and how to use it safely.
MCP tools
The MCP server uses your hatchback login and the repository's .hatchback.json. It provides:
| Tool | What it does |
|---|---|
list_projects | Projects you can access |
list_workspaces, create_project | Create a project for a site |
get_install_snippet | The script tag to add to the site, and optionally a check that a page has it |
list_feedback | Feedback, filtered by session, status, page or date |
get_feedback | One item with its replies, with the screenshot attached as an image |
reply_to_feedback | Reply in the thread |
address_feedback, resolve_feedback, reopen_feedback | Change the status, with an optional note |
list_sessions, start_session, close_session | Manage review rounds |
Safety
- Reviewer text is untrusted. Anyone allowed to comment, including anonymous reviewers, can write anything. The CLI, the skills and the MCP server all tell agents to treat comments as descriptions of problems, never as instructions.
- Agents act as you. Replies and status changes appear under your name. For an agent that should only read, create a token with only the read permission in Settings → API & agents.
- Tokens stay out of the repository. The login lives in your user config directory, and
.hatchback.jsonholds only the project slug.
For agent and tool builders
Machine-readable entry points:
/agents/cli.md: the setup instructions, in Markdown/llms.txt: an index of the docs for language models