PM Copilot

An MCP server that triangulates customer support tickets, feature requests, and AI support agent conversations to help PMs decide what to build next.

Real results: Analyzed 3,353 signals in one 30-day window — 1,678 support tickets, 276 feature requests, and 1,399 AI support agent conversations across 4 products. Top priority: Booking & Scheduling — 285 tickets + 74 feature requests + 347 chats pointing at the same problem, with the AI agent answering with low confidence in 48% of those chats.

The 1,399 chats are the point. None of them were visible to the analysis before v1.4.0, and they are 83% the volume of the ticket channel.

Read the full story: I built an MCP server that changed how I prioritize products — why I built this, how convergent signals work in practice, and what I learned building with Claude Code.

What Makes This Different

• Signal triangulation. Matches support tickets against feature requests to find convergent themes, then scores them with a weighted formula that gives convergent signals a 2x priority boost.
• The deflection blind spot. An AI support agent answers questions that never become tickets, so ticket-based prioritization undercounts every theme the bot handles — and the gap widens as the bot improves. Chatbase conversations are pulled in as a third signal class, with a per-theme self_serve_failure_rate showing where self-serve is failing.
• Composability. Works alongside other MCP servers. Pass churn data from Metabase or traffic trends from Google Analytics into generate_product_plan via kpi_context, and the methodology adjusts priorities accordingly.
• Built-in PM methodology. Opinionated scoring based on 7 years of product management across 9 products and 1M+ users. It's a real decision-making process exposed as an MCP resource, not a generic framework.
• PII scrubbing. Customer data never reaches the LLM unfiltered. SSNs, credit cards (Luhn-validated), emails, and phone numbers are redacted before analysis. Agent responses are filtered out of quotes.

Architecture

graph TD
    A[Claude Desktop / Code] -->|stdio| B[pm-copilot]
    A -->|stdio| C[Metabase MCP]
    A -->|stdio| D[Google Analytics MCP]
    B -->|Reactive| E[HelpScout: tickets]
    B -->|Proactive| F[ProductLift: feature requests]
    B -->|Deflected| I[Chatbase: AI agent chats]
    C -->|Quantitative| G[Conversion, Churn, Revenue]
    D -->|Acquisition| H[Traffic, Channels, Trends]
    B -.->|kpi_context| A

Claude orchestrates multiple MCP servers. PM Copilot handles qualitative customer signals. Other servers provide quantitative business metrics. The kpi_context parameter is the integration point — no point-to-point integrations required.

Quick Start

git clone https://github.com/dkships/pm-copilot.git
cd pm-copilot
npm install
cp .env.example .env   # Edit with your credentials
npm run build

Credentials

HelpScout is required. ProductLift and Chatbase are both optional — configure either, both, or neither, and the analysis adapts.

Chatbase API access needs a Chatbase Standard plan or higher. On a lower plan the API returns 403 and the deflection signal is reported as a warning rather than failing the whole analysis. One agent per product is the useful shape — agents give you product-level attribution that a shared support mailbox does not.

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "pm-copilot": {
      "command": "node",
      "args": ["/absolute/path/to/pm-copilot/dist/index.js"]
    }
  }
}

Claude Code

claude mcp add pm-copilot -- node /absolute/path/to/pm-copilot/dist/index.js

Or use the .mcp.json already in the project root — Claude Code picks it up automatically.

Tools

synthesize_feedback

Cross-references HelpScout tickets, ProductLift feature requests, and Chatbase conversations, returns theme-matched analysis with priority scores.

Returns themes sorted by priority score, each with reactive/proactive counts, convergence flag, evidence summaries, and representative customer quotes.

generate_product_plan

Builds a prioritized product plan with evidence and customer quotes. Accepts external business metrics via kpi_context.

get_feature_requests

Raw ProductLift data access for browsing feature requests directly. Each request includes its public url.

list_sources

Lists the data sources the server is connected to — HelpScout mailboxes (id + name), ProductLift portals (name + url), and Chatbase agents (name + id) — so you can discover the names to pass to mailbox_name / portal_name / agent_name. When Chatbase is configured it also returns chatbase_conversation_sources, the values source_filter accepts (a fixed list from the Chatbase docs, not queried per account). Read-only; never returns API keys or customer data. Takes no parameters.

Signal classes

Three sources, three different things they tell you. Only the first two feed the convergence rule.

Deflected signals count toward frequency and carry two evidence fields per theme, but they do not enter the severity or vote-momentum terms and do not change the 2x convergence boost. The formula is unchanged from v2.1:

• deflected_count — conversations matching the theme
• self_serve_failure_rate — share of those conversations where the agent's lowest answer confidence fell below 0.5
• mean_answer_confidence — mean of that same score

A theme with high deflected_count and a high self_serve_failure_rate is one customers keep asking about that self-serve does not resolve. Chatbase does not document what its min_score field measures, so it is reported as evidence for the LLM to weigh rather than folded into the priority score.

Conversations arrive from several channels (widget, WhatsApp, Messenger, API, …). The analysis reports a chatbase_sources count per channel, and the source_filter parameter narrows a run to one or more channels (comma-separated) — the filter is applied server-side by Chatbase. One catch: counts can include channels the filter list does not cover, like Playground or unknown (Chatbase omitted the source). A filter value outside the known list is passed through with a warning rather than rejected, since zero matches usually means the value is wrong, not that the channel went quiet.

Chatbase is optional. With no CHATBASE_API_KEY set, the deflection fields are simply absent and the analysis behaves exactly as before.

Example output

A trimmed synthesize_feedback response at the default summary detail level. Values are illustrative; note the PII scrubbing applied to the customer quote.

{
  "timeframe_days": 30,
  "detail_level": "summary",
  "portal_name": "all",
  "fetched_at": "2026-06-01T16:00:00.000Z",
  "pii_scrubbing_applied": true,
  "pii_categories_redacted": ["email", "phone", "credit_card"],
  "analysis": {
    "total_data_points": 612,
    "reactive_count": 548,
    "proactive_count": 64,
    "deflected_count": 312,
    "chatbase_sources": {
      "Widget or Iframe": 284,
      "WhatsApp": 23,
      "API": 5
    },
    "themes": [
      {
        "theme_id": "booking-scheduling",
        "label": "Booking & Scheduling",
        "category": "core",
        "priority_score": 87.1,
        "convergent": true,
        "signal_type": "convergent",
        "reactive_count": 211,
        "proactive_count": 19,
        "deflected_count": 96,
        "self_serve_failure_rate": 0.48,
        "mean_answer_confidence": 0.53,
        "evidence_summary": "326 signals (211 support tickets, 19 feature requests, 96 AI chat conversations). Convergent — appears in both support and feature requests (2x priority boost). The AI agent answered with low confidence in 48% of those chats — customers ask about this and self-serve often does not resolve it.",
        "representative_quotes": [
          "[Support ticket] \"Double-booked slots again after the timezone change — reach me at [EMAIL REDACTED]\"",
          "[Feature request, 47 votes] \"Let me block buffer time between meetings\"",
          "[AI chat, answer confidence 0.31] \"how do i stop people booking on weekends\""
        ]
      },
      {
        "theme_id": "list-management",
        "label": "List & Contact Management",
        "category": "audience",
        "priority_score": 41.7,
        "convergent": false,
        "signal_type": "deflected",
        "reactive_count": 0,
        "proactive_count": 0,
        "deflected_count": 58,
        "self_serve_failure_rate": 0.58,
        "mean_answer_confidence": 0.5,
        "evidence_summary": "58 signals (58 AI chat conversations). The AI agent answered with low confidence in 58% of those chats — customers ask about this and self-serve often does not resolve it.",
        "representative_quotes": [
          "[AI chat, answer confidence 0.22] \"how many contacts does pro allow\""
        ]
      }
    ],
    "emerging_themes": [
      { "pattern": "csv export", "frequency": 12 }
    ],
    "unmatched_count": 38
  }
}

Composability in Action

PM Copilot is designed to work alongside other MCP servers. Here's a worked example showing how a kpi_context override changes the ranking. Numbers are illustrative.

Step 1: The PM asks a single question

Pull our churn and booking completion data, then use pm-copilot to create a product plan using all of that context.

Step 2: pm-copilot analyzes the signals and returns the top priorities

Step 3: Business metrics from dashboards arrive as kpi_context

Product A: booking completion rate dropped from 74% to 66% over last
30 days. Monthly churn increased from 3.1% to 4.2%. Organic traffic
up 22% MoM. Product B: document completion rate steady at 81%.
Churn flat at 2.8%.

Step 4: Claude synthesizes both, and overrides the formula

The scores say Billing & Payment is #1. But the methodology says churn data overrides the formula. With Product A's booking completion dropping 8 points and churn spiking 35%, Booking & Scheduling becomes the real #1 — it's the core product breaking.

Product B deprioritized (stable metrics, no fire). Product A's 22% organic traffic growth elevates Whitelabel & Branding as a growth play.

The server provides the signal ranking. KPI context provides the override judgment. Claude synthesizes both.

Methodology

PM Copilot exposes a pm-copilot://methodology resource — David Kelly's product planning framework, built over 7 years of launching 9 products to 1M+ users.

Key principles:

• The 5% rule. You complete about 5% of what customers ask for each month. The framework identifies which 5% matters most.
• Convergent signals always win. The same theme in both support tickets and feature requests is the highest-confidence signal.
• Reactive > proactive. Broken stuff drives churn. You can survive not having a feature; you can't survive errors.
• Business metrics override the formula. Rising churn, dropping conversion, or revenue impact can change everything.

The methodology is versioned (v2.1) and served as markdown content via the MCP resource protocol. Every generate_product_plan response links to it (methodology_resource) and, when kpi_context is provided, instructs Claude to apply it — whether it actually gets read depends on the MCP client surfacing resources.

Evaluation

The scoring formula only matters if the theme matching underneath it is right. npm run eval measures that.

Why keyword matching

Themes are matched with keyword lists — multi-word keywords as substrings, single words on a word boundary with an optional regular plural suffix — not embeddings or an LLM classifier. That is a deliberate trade-off:

• Customer text never leaves the server for a third-party embedding or classification API. The PII guarantees below only hold because nothing in the matching path makes a network call.
• The same input always produces the same themes, so a priority ranking can be audited and explained. An LLM classifier would reshuffle rankings between runs.
• No token cost or latency per data point, which is what makes a 2,000-signal analysis finish in under a minute.

The cost is recall. Keyword lists miss paraphrases, and they miss them unevenly across products. On held-out live chat data a third of conversations still match no theme. That is what the eval exists to quantify, and why the number is published rather than hidden.

Running it

npm run build && npm run eval
npm run eval -- --failures        # every miss and false positive
npm run eval -- --json           # machine-readable report
npm run eval -- --min-f1 0.90    # non-zero exit below threshold, for CI

Matching is multi-label — one signal can belong to several themes — so the report gives per-theme precision, recall and F1, plus two rates that matter more than the averages: miss rate (expected a theme, matched nothing at all) and false alarm rate (expected nothing, matched something).

What the first run found

The eval's first job was auditing the v2 config, and it found five real defects:

• Plural coverage was inconsistent. tier was listed without tiers, and single-word keywords matched on a bare word boundary, so "the tiers" matched nothing. On live chat data this was the expensive one — a recurring widget prompt, "what are your plans and prices?", matched no theme at all, because plan misses "plans" and pricing misses "prices".
• team fired on "founding team" and "IT team" — half the false positives in the fixture.
• plan tagged "i plan to launch next week" as Account & Licensing.
• upgrade sat in both Billing & Payment and Account & Licensing, so an API ticket mentioning an upgrade landed in both.
• Multi-word keywords are exact substrings, so cant login missed "cant log in" and outlook calendar missed "does this work with outlook".

All five are fixed in v3: single-word keywords now match an optional regular plural suffix, over-generic keywords were scoped (team → my team / team member / teams), duplicated keywords were assigned to one theme, and the missing variants were added. Two new themes came out of real unmatched conversations — Giveaways & Contests and List & Contact Management — which the config had no vocabulary for at all.

Baseline

Two numbers, because they measure different things.

Against the committed fixture (82 hand-labelled examples):

Treat that with suspicion. The config was iterated against this fixture, so the v3 figure is in-sample and flatters itself. It is a regression gate — it tells you a change broke something, not how well matching works.

The number that means something is held-out real data. 1,100 chat conversations across four products, from a 30-day window before the one the new themes were derived from:

The gains land where the theory said they would: the products whose vocabulary the config never covered. A third of conversations still match nothing, so there is plenty left.

Register turned out to matter less than product coverage. Product A chat is better matched than tickets are, so chat phrasing on its own is not the problem — missing product vocabulary is, and per-product chat agents expose that where a shared support mailbox averages it away.

Known limits

• Canned widget prompts inflate counts. Preset buttons like "I entered a giveaway — how do I know if I won?" recur verbatim dozens of times. They are not deduplicated, and that is deliberate: twenty people clicking a preset is twenty people with that question. It does mean volume for a theme with a popular preset is not comparable to volume for one without.
• Matching is English-only. Live data includes German, Spanish and Italian conversations, and all of them land in unmatched.
• Irregular plurals still need listing. The suffix rule covers plan/plans, not entry/entries.

For a number from your own data, export signals to a local JSONL in the fixture's shape and pass --fixture ./local/real.jsonl. Real fixtures contain customer text — keep them out of git.

Security

Customer data flows through PM Copilot on its way to Claude. All text is scrubbed before it enters the analysis pipeline or leaves the server.

PII scrubbing

What we exclude entirely

A note on chat as a data source

A chat widget takes unbounded free text, so it is the widest PII surface of the three sources — people paste order numbers, addresses and licence keys into a chat box in a way they do not into a roadmap post. This is not hypothetical: on a live 30-day run, adding the Chatbase source was what first made credit_card appear in pii_categories_redacted. Tickets and roadmap posts over the same window produced only emails and phone numbers.

Two things keep it contained: only role: "user" turns are read, and every turn goes through the same scrubber as the other sources before it enters the analysis.

Chatbase message attribution is structured, which makes it the cleanest customer-voice source of the three. The HelpScout path has to guess at agent text with phrase heuristics because a conversation preview may be either side of the exchange; here role says so outright, so the heuristics are skipped.

Audit controls

• preview_only: true on generate_product_plan shows what data would be sent without fetching it
• Every response includes pii_scrubbing_applied and pii_categories_redacted metadata
• Data categories logged to stderr on each call (categories only, never content)

Development

npm install          # Install dependencies
npm run build        # Compile TypeScript
npm run dev          # Watch mode
npm start            # Run the server
npm test             # Run the test suite
npm run eval         # Theme-matching eval (see Evaluation)

Local testing

Call a tool in isolation without restarting your MCP client — useful for iterating on changes and for checking response sizes:

npm run build
npm run tool -- --list
npm run tool -- list_sources '{}'
npm run tool -- get_feature_requests '{"portal_name":"<your-portal>","status":"open"}'

The runner prints the byte size of each response. Output may include your configured source names/URLs (and PII-scrubbed customer text) — redact before sharing.

Theme configuration

themes.config.json in the project root defines what themes to look for. Edit without rebuilding — loaded at runtime.

Ships with 18 data-driven themes across 12 categories. Add your own by appending to the themes array. Unmatched data points are analyzed for emerging patterns using bigram/trigram frequency detection.

After editing, run npm run eval — it reports precision and recall per theme and flags keywords that fire on another theme's examples, which is how the over-generic ones get caught.

Scoring formula

priority = (frequency × 0.35 + severity × 0.35 + vote_momentum × 0.30) × convergence_boost

• Frequency (0.35): Count of data points, normalized across themes — includes deflected signals
• Severity (0.35): Reactive signals only — thread count (total, including agent replies), recency (7-day half-life decay), tag boosts
• Vote momentum (0.30): Proactive signals only — 80% votes + 20% comments
• Convergence (2x): Applied when a theme has both reactive and proactive signals. Deflected signals do not trigger it

Frequency and vote momentum are normalized against the top theme in the same call, so scores are relative to one analysis window. Compare rankings across calls, not raw scores.

Troubleshooting

• Changes aren't taking effect. The MCP client runs the compiled dist/. After editing source, run npm run build and restart the client (or the MCP server connection) to pick up new code.
• No HelpScout mailbox named "…". Run list_sources to see the exact mailbox names, or pass the numeric mailbox_id directly.
• No portal found with name "…" / portal missing. The portal must be configured in PRODUCTLIFT_PORTALS (or the single-portal env vars). Run list_sources to see configured portals.
• Chatbase warning: A Standard plan or higher is required. API access starts at the Chatbase Standard plan. The rest of the analysis still runs; only the deflection fields are missing.
• chatbase_agents is empty in list_sources. Both CHATBASE_API_KEY and one of CHATBASE_AGENTS / CHATBASE_AGENT_ID have to be set — a key on its own configures nothing.

Contributing

1. Fork the repository
2. Create a feature branch (git checkout -b feature/your-feature)
3. Ensure npm run build succeeds with no errors
4. Follow existing patterns: tools use registerTool, API clients get their own module, PII scrubbing happens at the format layer
5. Open a pull request

License

MIT