Wellness Planner

A local MCP agent that queries personal health data and provides energy-aware task scheduling.

Data Note

This project uses simulated data. Health data is seeded from data/seed_db.py into a SQLite database. There is no Apple Health integration — real health data is not imported or synced.

Running the Code

Prerequisites

• Python 3.14+

• uv for dependency management

Standalone Agent (no MCP server required)

Run the Plan-and-Execute agent loop locally:

uv run python mcp_server/agent.py [YYYY-MM-DD]

• Uses yesterday's date if no date is given.

• Reads from data/health.db and data/todo.json.

• Prints a daily brief: sleep, activity, heart rate, readiness score, and proposed schedule.

MCP Server (for Cursor)

The MCP server is spawned by Cursor when needed — you do not start it manually in a separate terminal.

1. Configure Cursor to use the local MCP server (e.g. .cursor/mcp.json):

{
  "mcpServers": {
    "wellness-planner": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/wellness_planner", "python", "mcp_server/server.py"]
    }
  }
}

2. Replace /path/to/wellness_planner with your actual project path.

3. Cursor will spawn the server and communicate over stdio.

Other Commands

MCP Tools

When the server is connected, these tools are available:

• get_health_summary — Aggregated sleep, activity, and heart rate for a date

• calculate_readiness_score — 1–10 readiness score for task timing

• query_raw_logs — Run read-only SQL against the health DB

• get_tasks — Load tasks from todo.json

• propose_schedule — Energy-aware schedule based on readiness and tasks

• get_data_dictionary — Schema introspection: column names, types, and sample values

• run_analysis — Execute a pandas/sqlite analysis script locally; returns stdout

• generate_chart — Produce a self-contained Observable Plot HTML chart

• get_insights — Retrieve previously saved findings from the Fact Store

• save_insight — Persist a discovered insight so it isn't re-computed next session

Testing

There are two layers to test: the skills directly, and the MCP tools through Cursor chat.

1. Test skills directly (fast, no Cursor needed)

Phase 1 — Sandbox execution:

uv run python -c "
from skills.sandbox import run_python_analysis
r = run_python_analysis('''
df = pd.read_sql('SELECT date, total_hours FROM sleep_logs ORDER BY date DESC LIMIT 7', __import__('sqlite3').connect(DB_PATH))
print(df.to_string(index=False))
''')
print(r['output'])
"

Phase 2 — Schema discovery:

uv run python -c "
from skills.schema import get_data_dictionary
import json
print(json.dumps(get_data_dictionary(), indent=2))
"

Phase 3 — Chart generation:

uv run python -c "
import sqlite3
from skills.visualization import generate_chart
rows = sqlite3.connect('data/health.db').execute('SELECT date, total_hours FROM sleep_logs ORDER BY date').fetchall()
r = generate_chart([{'date': r[0], 'total_hours': r[1]} for r in rows], 'Sleep Trend', 'date', 'total_hours')
print(r)
"

Then open the url value in a browser to see the chart.

Phase 4 — Fact Store:

uv run python -c "
from skills.memory import save_insight, get_insights, clear_insights
save_insight('test_key', 'test value', 'manual test')
print(get_insights())
clear_insights()
"

2. Test end-to-end through Cursor (the real agentic loop)

Ask the agent questions in chat and watch the MCP tool calls fire in sequence:

• Schema discovery: "What tables and columns are in the health database?"

• Analysis: "What's the correlation between my step count and sleep quality over the last 30 days?"

  • Should trigger: get_insights → get_data_dictionary → run_analysis → save_insight

• Chart: "Show me my resting heart rate trend as a chart."

  • Should trigger: run_analysis → generate_chart → returns a file path

• Memory: "What do you already know about my health patterns?"

  • Should trigger: get_insights and return stored findings without re-running anything

3. Standalone agent CLI

uv run python mcp_server/agent.py 2026-02-18

Tests the non-MCP path (summarizer + readiness + scheduling) and confirms nothing broke during the Phase 1–4 additions.