An MCP server for website feedback
What the Feedback DevTools MCP server gives a coding agent, how to connect it, and which tools are available on the hosted server and the npm package.
Why a coding agent needs the feedback, not a summary of it
A coding agent is only as good as the context it starts with. Pasting a screenshot and a sentence into a chat loses the console error, the failed request and, most of all, the file that renders the element. The Feedback DevTools MCP server hands the agent the report as data instead: the reviewer's words, the element, the component and source file:line from the React SDK, and the technical evidence. The agent starts in the right file with the reason for the change, which is the difference between a guess and a fix.
What the agent can do
Reading tools let the agent list open feedback, fetch one item with its full context and a ready-made prompt, list projects, and see analytics and classification counts. Writing tools let it claim an item so two agents never take the same one, update status, priority or category, ask the reviewer a follow-up question, share an item as a public link, and mark it resolved with the pull request URL it opened. Issue tools create a GitHub issue or a Linear issue whose body carries the source-mapped context.
Agent-run tools start a hosted fix for an item, report the run's status and whether it opened a pull request, and approve the run. Opening the pull request from a hosted run needs Autopilot on a Pro project, and merging always stays with a person. The server also offers a separate read-only endpoint for agents that should look at feedback but never change it.
Hosted endpoint or local npm package
The hosted MCP endpoint is scoped to one project with its SDK token and authenticated with your dashboard session, so there is nothing to install. It carries the complete tool list. The local server runs over stdio with npx, takes the API URL, SDK token and an auth token as environment variables, and suits teams that want the server to run inside their own machine.
The npm package is a version behind the hosted server. These tools are available on the hosted endpoint only until version 0.2.0 of feedback-devtools-mcp is published: claim_feedback, create_github_issue, create_tracker_issue, get_cluster, get_next_feedback, list_clusters. Everything else, including listing and reading feedback, claiming, updating and resolving items, works the same in both. The setup pages for Cursor, Claude and VS Code show the exact configuration for each editor.
Frequently asked questions
What is the Feedback DevTools MCP server?
A Model Context Protocol server that gives coding agents access to your website feedback: open items with screenshots, console and network logs and source file:line, plus tools to claim, update and resolve items and to create GitHub or Linear issues.
Which editors and agents does it work with?
Any MCP client. Setup guides cover Cursor, Claude Code and Claude Desktop, and VS Code with Copilot.
Can the MCP server open pull requests?
Your own agent opens the pull request from your checkout and records its URL with mark_feedback_resolved. On Pro projects with Autopilot, start_agent_run starts a hosted fix that opens the pull request itself. Neither merges.
Is the MCP server free?
Yes. It works on the Free plan. Hosted agent runs through Autopilot need Pro, at $9/mo per project.
Keep reading
Try it on your own site
Feedback DevTools is free to start: unlimited feedback, shareable review links and no client accounts.