If an AI agent is already writing code in your editor, it should be able to read what your users are asking for too, without you copying feedback into a prompt by hand. Every Upstep project now ships a built-in MCP server that does exactly that.
What it can do
- get_project_overview and list_feedback, so an agent can see what's open and what's winning by votes
- create_feedback, to file a task, defaulting to Dev-only so it never shows up in the public widget
- update_feedback, to move a card across your board by column name
- add_comment, to reply to the users who asked for it
- create_board, for a workspace that's entirely the agent's own
Connecting it takes either one Claude Code command or a short Codex setup, both scoped to a single project by a private MCP key:
claude mcp add --transport http upstep https://upstep.dev/api/mcp \
--header "Authorization: Bearer YOUR_MCP_KEY"
export UPSTEP_MCP_KEY="YOUR_MCP_KEY"
codex mcp add upstep --url https://upstep.dev/api/mcp \
--bearer-token-env-var UPSTEP_MCP_KEYDev-only, by default
The part we spent the most time on isn't a tool call, it's a default. When an agent files a task through the MCP server, it lands as Dev-only unless told otherwise, hidden from the public widget. Your users see their feature requests and bug reports. They don't see your agent's internal refactor list. It can even run its own board entirely apart from the one you show visitors.
Get connected
Open any project, head to the new MCP tab, and copy the Claude Code command, the Codex setup, or the JSON config for Cursor and other clients. The setup guide walks through the same thing if you're wiring up the widget for the first time.