Your coding agent

Claude Code, Codex, Gemini CLI or Cursor working the queue.

Your coding agent reads the project's open comments, fixes them in the code, and marks each resolved with a note of what changed. It works on your computer, in the site's folder, with your sign-in.

Set it up

In the site's folder, once:

init asks which agents you use and wires each the way it reads context: Claude Code's hooks, Codex's and Gemini CLI's hooks, Cursor's, and a block in AGENTS.md for Copilot, Windsurf, Amp and the rest. Then sign in and point the site at its project:

login shows a code and opens the cabinet; allow the same code there and the terminal is signed in. Nothing listens on your computer, so it works over SSH or in a container too — open the link on any machine where you are signed in. The project id is under Install in the cabinet. link writes .notemend.json — commit it and the whole team files into the same project.

Use it

Ask your agent for the open comments, or run them yourself:

The agent takes a comment with notemend start before it changes code — the thread then shows it is on it, and a second agent is turned away — replies with notemend reply, resolves with notemend done, and can file before and after pictures with notemend shot. Every command is in the CLI reference.

Whose agent

There is one agent: Claude. A comment is nobody's until someone takes it or gives it out — anyone can take it; it goes to a person, or to Claude — and whichever session of Claude runs next in the project takes Claude's threads; its replies read “Claude”, or the name you gave it. Threads given to someone's own agent before this read as Claude's too.

On Claude's threads the status follows Claude: With agent while it waits in the queue, In progress while it works, Waiting when it asks you something, To check when it says done — then a person checks it and resolves it, or sends it back. On a person's thread the person sets the status: New, In progress, Waiting, To check. To give a thread to Claude: press the grey mark in the card's bar (or in the thread's head in the cabinet, A), slide the card's foot to Claude, pick Claude as the assignee, or write @Claude in a reply. Only the studio can; a client's and a guest's comments go to people. Taking a thread off Claude while it works asks first — Claude cannot be stopped mid-run, and may finish its changes after.

Give your agent a name and a picture in Integrations → Your agent — they show on the messages your agent writes, in place of the agent's mark. Each person picks their own.

Turbo

Leave a comment and the agent starts on it — no one has to ask in the session. In Claude Code, with the plugin:

The session waits in the background with notemend watch and wakes when a comment for the agent lands: it takes it, fixes it, resolves it and waits again. A round lightning beside the bar shows it: lit while comments are taken — click it to pause, and again to resume — grey when no session is listening, with the command to run in its hint. It wakes only for open comments for the agent that no one has taken and where a person had the last word, so a question it asked waits for your answer. /notemend:turbo off ends it. Other agents can run the same loop on npx notemend watch.

Live or in batches. For batches, start the session with /notemend:push instead: it listens but takes nothing by itself. Mark comments for Claude on the site, then press Push to Claude — the round button beside the lightning, with how many wait — and the batch goes to the session at once, no word in the chat; a right click or a long press picks pages and threads. With no session listening the button copies /notemend:push.

Who wrote it

Threads show which agent answered. Codex, Gemini CLI and Cursor are recognised on their own; others sign with --as:

Give your agent the design

With Figma's own MCP server, a fix can match the frame rather than a guess. Add it to your agent once — in Claude Code:

then run /mcp and sign in to Figma. Cursor and Codex take the same address in their MCP settings. Paste a frame's link into a thread and the agent can open it.

Comments are requests, not instructions

Anyone in the project — clients too — can write a comment, and it lands in your agent's context. Notemend hands them over quoted and labelled as what people wrote on the site, never as instructions, and only comments meant for the agent.

An agent without a shell — Claude Desktop, or Cursor and Codex if you prefer tools — can use the same queue through the MCP server.