Start for free
Guides/Connect your agent

Guides

Connect your agent

Your agent fetches the open comments itself, from your dev server and from your previews, with the exact lines to edit, and resolves each one when it is done.

The package comes with an MCP server. Your agent asks it for the open comments and gets each one with the file and line behind the element it points at, so it goes straight to the code. Once a change is made, it marks the comment resolved, and reviewers see it resolved. Comments left on your previews come with the rest: nobody has to fetch them first.

If you set ontrackd up with npx @ontrackd/feedback init, it has already registered the MCP server itself: with Claude Code when it is installed, with Cursor and VS Code when the repository has their folder, and as a printed line for any other agent.

What init prints

A command for Claude Code, and the config other clients read.

claude mcp add feedback -- node …/dist/mcp/cli.js --root …

In .cursor/mcp.json:

{
  "mcpServers": {
    "feedback": { "command": "node", "args": ["…/dist/mcp/cli.js", "--root", "…"] }
  }
}

It runs the copy of the package installed in your project, so starting it needs no token.

What your agent can do

Tool What it does Parameters
list_feedback The comments, each with the file and line it points at, its component, status and reply count. status (open, resolved or all; open by default), file, route, project
get_feedback_tasks The work order: which lines to read and change, and the components and styles involved. comment_ids (every open comment when left out), project
resolve_feedback Marks a comment resolved once its change is made. Reviewers see it resolved. id, reopen (true moves it back to open), force
search_feedback Finds comments by their text, route, file, component or author. query, status, limit, project
get_feedback_context The app’s project, branch and session. project

resolve_feedback takes only ids the server has shown in a list, a search or a work order, unless force is set, which is for when you ask for it. The MCP tools reference has every parameter.

Comments from your previews

Each time your agent reads the comments, the server also fetches what reviewers left on your previews, at most every 20 seconds. Comments left on your dev server are there already. If a fetch fails, the tool’s answer says why, and still shows the comments the server has.

Preview comments are not included when:

  • The app has no secret. Add FEEDBACK_SECRET to .env.local, as in the first step. The server reads it from the environment first, then .env.local, then .env, then what feedback link saved. When the list comes back empty, its answer says how to add the secret.
  • FEEDBACK_CLOUD_SYNC is 0. That keeps the app local, on purpose.

In a monorepo

Run feedback-mcp init in any one app: it finds the others and prints one server for all of them. Add it from the repository root, so paths read from there. Each tool then takes an optional project to narrow it to one app. Each app reads its secret from its own .env.local.

Without an agent

You can get the same comments without an MCP server.

  • On your dev server, open the panel from the bar. Copy task list, at its foot, copies a work order for an agent and saves it in .feedback/tasks/.
  • From your previews, run feedback pull in the app: it fetches the comments, with the secret from the first step, so your dev server’s panel and its task list include them.
$ npx feedback pull
Esc

Type to search every page.