Rank what to build next with an open-source MCP server that scores support tickets and feature requests, then lets Claude write the plan

Every month the author had to read more than 200 pieces of customer feedback across three products, scattered over a support inbox, a feature-request board and analytics. Copying it into a document and asking Claude to summarise it by hand was slow and gave different answers each time.

1. Connect the server to Help Scout (support tickets) and ProductLift (feature requests) through their APIs. 2. Pull the signals and convert them to one common format. 3. Remove personal data before anything reaches the model: card numbers (checked with the Luhn rule), social security numbers, emails and phone numbers. 4. Match each signal to a list of editable themes and score each theme: frequency 35%, severity 35%, vote momentum 30%, doubled when the same theme shows up in both tickets and requests. A recent spike in tickets counts more than a steady pile of votes, and many votes with no tickets is read as a want rather than a need. 5. Paste current business numbers (for example a drop in booking completion or a rise in churn) into a free-text KPI field, which can override the formula. 6. Claude turns the scored JSON into a prioritised product plan. The author makes the final call and still speaks with two to four customers a month. The server was built with Claude Code in about 20 hours and is published on GitHub (dkships/pm-copilot).

In one live run the author reports 10,424 signals (10,190 tickets, 234 requests) across three products processed in 55 seconds into 16 themes, 15 of which appeared in both sources. Without business context, billing and payment ranked first; after adding a fall in booking completion (74% to 66%) and a churn rise (3.1% to 4.2%), booking and scheduling moved to first. The article does not report what was built afterwards or whether churn improved.

Built for a three-product software company with a small team and a large user base (one product cites 350,000+ users), not a solo founder, so the volume is much higher than most readers will have. Only the Help Scout and ProductLift connectors exist; metrics come in as free text. The scoring weights reflect the author's own judgement. Large outputs exceed the model's context and get truncated. Claude Code badly underestimated the data size (167 conversations estimated, 2,134 actual). The Claude model version is not named. No result after acting on the plan is reported. Firsthand, single run, self-reported numbers.

Help Scout and ProductLift accounts with API access, Claude with MCP support, and enough Node.js comfort to install and configure a TypeScript MCP server.

Edit the theme list and the weights before trusting the ranking; the author says they are meant to be changed for each business.

Here is a scored list of customer themes from our support tickets and feature requests, plus this month's key numbers: [paste]. Rank the top five things to work on next. For each, say which signals support it, whether it looks like a need or a want, and what you would check with two customers before building it.

• dmkthinks.org (David Kelly) ↗ Firsthand write-up, published February 26, 2026, main source
• GitHub (dkships) ↗ Code repository, publication date unknown

Source published: February 26, 2026. Last reviewed here: October 30, 2026. Spot a mistake? Tell us.