Pm Agent

An MCP server that gives Claude the tools to help a product manager make sprint decisions grounded in real data. Built for DevPulse's PM Asha: four tools that answer the questions she spends 60% of her week answering manually.

Tools

Tool	Purpose
prioritize_backlog	Score and rank backlog items by RICE, customer signal, or a combined method. Flags stale, unestimated, blocked, and anomalous items.
analyze_feedback	Extract themes from customer feedback with ARR weighting and bias warnings (over-represented customers, churned signal, segment skew).
assess_capacity	Compute per-engineer available capacity for the sprint, accounting for allocation %, PTO, and carry-over work.
map_dependencies	Trace dependency chains, detect cycles, and flag external blockers and long chains for a set of backlog items.

Prerequisites

• Python 3.10+
• uv (recommended) or pip

Setup

cd mcp_starter
python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -r requirements.txt

Place the five data files in ./data/:

data/
  product_backlog.json
  customer_feedback.json
  team_roster.json
  dependency_map.json
  sprint_history.json

Running

python server.py

The server uses stdio transport (what Claude Desktop and Claude Code expect). It will not print anything to stdout on startup — that's normal.

DATA PATH CONTRACT

The server reads its dataset from the PM_AGENT_DATA environment variable, falling back to ./data for local development:

DATA_DIR = Path(os.environ.get("PM_AGENT_DATA", Path(__file__).parent / "data"))

At grading time the evaluation harness mounts a different dataset at PM_AGENT_DATA. No IDs, names, or numbers are hardcoded — all tools compute from whatever is mounted.

Connecting to Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "pm-agent": {
      "command": "python",
      "args": ["/absolute/path/to/mcp_starter/server.py"],
      "env": { "PM_AGENT_DATA": "/absolute/path/to/data" }
    }
  }
}

Use absolute paths. Restart Claude Desktop after editing.

Connecting via Claude Code

claude mcp add pm-agent python /absolute/path/to/mcp_starter/server.py

Tool Reference

prioritize_backlog

Score and rank backlog items by one of three methods.

Parameter	Type	Default	Description
method	string	"combined"	"rice", "customer_signal", or "combined"
squad	string	null	Filter to "platform" or "growth"
status_filter	list	null	e.g. ["planned", "proposed"]
include_dependency_check	bool	true	Flag and penalize blocked items
limit	int	20	Max items returned

Flags: STALE, UNESTIMATED, NO_CUSTOMER_SIGNAL, BLOCKED, EXECUTIVE_PRIORITY_ANOMALY, LOW_CONFIDENCE, DUPLICATE_TITLE

---

analyze_feedback

Extract themes from customer feedback with bias detection.

Parameter	Type	Default	Description
theme_limit	int	5	Number of top themes
group_by	string	"theme"	"theme" or "customer"
customer_status	string	null	Filter: "active", "churned", "trial"
customer_tier	string	null	Filter: "enterprise", "mid_market", "startup"

Bias warnings: OVER_REPRESENTED_CUSTOMER, CHURNED_CUSTOMER_SIGNAL, SEGMENT_SKEW, ARR_CONCENTRATION

---

assess_capacity

Per-engineer sprint capacity with three tiers: total → effective (after allocation/PTO) → available (after carry-over).

Parameter	Type	Default	Description
squad	string	null	Filter to "platform" or "growth"
required_skills	list	null	e.g. ["backend", "security"]

Formula: effective = 21 × (allocation%/100) × ((10 - pto_days)/10) · available = effective - carry_over_points

---

map_dependencies

Trace dependency chains and surface risks.

Parameter	Type	Default	Description
item_ids	list	required	e.g. ["BP-112", "BP-117"]
max_depth	int	3	Hops to follow
include_soft	bool	true	Include non-blocking soft deps

Risk flags: CYCLE, EXTERNAL_NO_ETA, EXTERNAL_WITH_ETA, LONG_CHAIN

Repo Structure

mcp_starter/
├── server.py            # MCP entry point
├── requirements.txt
├── olympics.json        # run contract
├── README.md
├── TDL.md               # Technical Decision Log
├── tools/
│   ├── __init__.py
│   ├── prioritize_backlog.py
│   ├── analyze_feedback.py
│   ├── assess_capacity.py
│   └── map_dependencies.py
└── data/                # sample data for local dev (not committed)