If you spend most of the day in Claude Code or Cursor, checking what customers asked for means another tab, another login, and a context switch at the moment you least want one. CatchLoop has an MCP server, so the agent you already work with can read your board and update it from where you are.
This post shows how to connect it, four things it is useful for, and what it does not do yet.
What the MCP server gives your agent
The Model Context Protocol is a standard way for an AI client to call tools on another service. CatchLoop hosts the server, so there is nothing to install. There are five tools, all free, writes included. A sixth, publishing a changelog entry from the agent, will come with Pro, which is not open yet, and only appears on a workspace where it is on: for now the agent drafts the entry and you publish it in the dashboard.
| Tool | What it does |
|---|---|
list_requests | The most-wanted requests, ranked by the votes of your customers' accounts. It can narrow to one or several statuses and return up to 100 at a time. |
create_request | Adds one request, with a title and a description. |
import_requests | Adds up to 50 requests in one call, with their past vote totals, original dates and statuses. |
set_request_status | Moves a request to another status. It emails nobody. |
create_changelog_draft | Drafts a changelog entry linking the requests it ships, and says how many people publishing it would email. |
publish_changelog_entry | Publishes that draft: its requests move to Shipped and their people are emailed, once. It takes back the count you confirmed. Part of Pro, off by default, and not offered until then. |
Each request comes back with its id, title, status and two vote numbers: the votes of signed-in accounts, and the votes imported from another tool. The public board shows their sum.
Connect it
Your workspace key is on the Integration page of the dashboard, under MCP server. Keep it in an environment variable rather than in a file you commit:
export CATCHLOOP_API_KEY="your workspace key"Claude Code
One command, run once:
claude mcp add --transport http catchloop https://mcp.catchloop.co \
--header "Authorization: Bearer $CATCHLOOP_API_KEY"Cursor
One entry in the project's MCP file. Cursor reads the variable itself:
{
"mcpServers": {
"catchloop": {
"url": "https://mcp.catchloop.co",
"headers": { "Authorization": "Bearer ${env:CATCHLOOP_API_KEY}" }
}
}
}VS Code and Devin Desktop (Windsurf) take a similar entry; the MCP docs have both.
Their custom connectors sign in with OAuth, which CatchLoop does not offer yet. If your organization has Anthropic's request headers beta, you can add a custom connector with the server URL, choose No sign-in, and add an authorization header whose value is Bearer, a space, then your key.
Four things to ask it
What should we build next?
The question the board exists to answer, one prompt away:
List our ten most-wanted requests that are New or Planned, and tell me which ones look related.The ranking is by account votes, because an account is a person who can be emailed when the request ships. The agent only sees titles and numbers, not descriptions or comments, so treat its grouping as a first pass, not a verdict.
Log feedback where you read it
A customer explains a problem in a support thread or a call summary. Rather than retyping it on the board, paste it to the agent:
Create a request from this ticket. Keep the customer's wording for the problem, leave out their name.The request lands in New on your public board, where other customers can find it and vote. Read it before you move on: it is public the moment it is created.
Bring a backlog in
If your requests live in a spreadsheet or another tool's export, the agent can carry them over, fifty per call, with their vote totals, dates and statuses, so the board starts ranked instead of flat. Importing emails nobody. The same thing works without an agent: the dashboard takes a pasted list or CSV, described in the import docs.
Keep statuses in step with the code
When you start on something, say so where your customers can see it:
Find the dark mode request and move it to In progress.The agent looks the request up with list_requests, then calls set_request_status with its id. Nobody is emailed. That is deliberate: in CatchLoop, customers hear once, when you publish the changelog entry that ships their request, and that step always has a person deciding. The agent can draft the entry with create_changelog_draft and tell you how many people it would email; you read it and publish it from the dashboard's Changelog page.
Try it on your own board
A free workspace comes with its MCP key. No volume caps, writes included.
What it does not do
- It does not publish for you, yet. The agent drafts, you publish in the dashboard. Publishing from the agent will come with Pro, and even then it takes back the recipient count you confirmed, and refuses if it changed since the draft.
- It does not read or write comments, and it deletes nothing. Five tools, six with Pro, no more.
- It has no OAuth yet, which is why claude.ai and Claude Desktop need the beta described above.
- It is rate limited per workspace key. The numbers are on the rate limits page.
Look after the key
The key is a workspace credential. Anything your agent can do with it, anyone holding it can do: read your requests, add new ones, move statuses, draft changelog entries, and, on a workspace where publishing from MCP is on, publish one that emails your customers. Keep it in the environment, out of committed files and out of screenshots. Every call is scoped to the workspace the key belongs to, so it can never reach another team's board.
The full reference, with every parameter and status value, is in the MCP docs, and the MCP page has the snippets for each client.