Fitness Agent MCP

A zero-dependency Express API that acts as a universal fitness intelligence layer for AI assistants. Claude, ChatGPT, and GitHub Copilot all share the same state, gamification engine, and data. Claude and Copilot via the built-in MCP server, ChatGPT via OpenAPI Custom Actions.

No OpenAI key. No database server. One deployed URL.

---

Table of Contents

1. What This Does
2. Architecture
3. Local Setup
4. Deploy to Railway
5. Connecting AI Clients
6. MCP Tools
7. Security Notes
8. Environment Variables

---

1. What This Does

One server, three AI surfaces:

AI Client	Protocol	How it connects
Claude Desktop / claude.ai	MCP (StreamableHTTP)	Calls /api/mcp with JSON-RPC
GitHub Copilot Chat	MCP (SSE)	Reads .vscode/mcp.json, calls /api/mcp
ChatGPT Custom GPT	REST / OpenAPI	Reads /api/openapi.json, calls REST endpoints

Features: user profiles · workout + diet plans · calendar scheduling · gamification (XP / streaks / 12 achievements) · paginated history · export (JSON / CSV / interactive HTML) · cron reminders

Design principle: The server is a pure storage layer. The AI client (Claude, GPT-4, Copilot) reads tool descriptions and generates all structured data itself — diet plans, workout schedules, calendar events — then passes them as parameters. No server-side AI calls. No API keys required.

---

2. Architecture

Claude / Copilot          ChatGPT Custom GPT
    └─ MCP JSON-RPC ──────┐   └─ REST/OpenAPI ──┐
                          ▼                      ▼
              ┌──────────────────────────────────────┐
              │        Express API  (/api/*)         │
              │                                      │
              │  MCP tools (8)     REST routes       │
              │  get_state         /state            │
              │  save_state        /log-completion   │
              │  log_completion    /normalize        │
              │  normalize_user_input /generate-plan │
              │  generate_plan     /schedule-events  │
              │  schedule_events   /export           │
              │  get_history       /progress/history │
              │  export_report     /healthz          │
              │                                      │
              │  Gamification engine + node-cron     │
              └──────────────────┬───────────────────┘
                                 │
                                 ▼
                        ┌────────────────┐
                        │  SQLite (file) │
                        │                │
                        │ user_profiles  │
                        │ diet_plans     │
                        │ workout_plans  │
                        │ schedules      │
                        │ progress       │
                        └────────────────┘

Tables are created automatically on first boot via Drizzle migrations — no manual schema push needed.

---

3. Local Setup

Prerequisites

• Node.js 20+
• pnpm 9+

Install and run

git clone <your-repo-url>
cd fitness-agent-mcp-v2
pnpm install

Copy the example env file (optional — defaults work out of the box):

cp .env.example .env

Start the server:

pnpm --filter @workspace/api-server run dev

The API is at http://localhost:8080/api. SQLite database is created automatically at ./artifacts/api-server/data/fitness.db.

Verify

curl http://localhost:8080/api/healthz
# → {"status":"ok"}

Run tests

pnpm --filter @workspace/api-server test

---

4. Deploy to Railway

The railway.toml at the project root configures everything automatically.

Steps

1. Push this repo to GitHub
2. railway.app → New Project → Deploy from GitHub repo → select repo
3. Dashboard → Add Volume → Mount path: /data, Size: 1 GB
4. Service → Variables → add:

Key	Value
PORT	8080
DB_PATH	/data/fitness.db

5. Deploy — Railway auto-sets RAILWAY_PUBLIC_DOMAIN

Test

curl https://your-app.up.railway.app/api/healthz
# → {"status":"ok"}

List on Smithery (optional)

Smithery.ai is the MCP marketplace — users can find your server and connect it to Claude with one click.

npx smithery mcp publish "https://your-app.up.railway.app/api/mcp" \
  -n yourusername/fitness-agent

---

5. Connecting AI Clients

Once deployed, get your base URL (https://your-app.up.railway.app).

Claude — claude.ai web

1. Settings → Integrations → Add Custom Integration
2. Paste: https://your-app.up.railway.app/api/mcp
3. Done — syncs to Claude iOS app automatically

Claude Desktop (manual config)

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "fitness-agent": {
      "url": "https://your-app.up.railway.app/api/mcp",
      "transport": "streamable-http"
    }
  }
}

Restart Claude Desktop completely after saving.

Get the recommended system prompt:

curl https://your-app.up.railway.app/api/system-prompt
# → use the claude.prompt value

GitHub Copilot — VS Code

Update .vscode/mcp.json with your deployed URL:

{
  "servers": {
    "fitness-agent": {
      "url": "https://your-app.up.railway.app/api/mcp",
      "type": "sse"
    }
  }
}

.github/copilot-instructions.md is already in the repo — Copilot Chat reads it automatically for full API context.

ChatGPT — Custom GPT Actions

1. chat.openai.com → profile → My GPTs → Create → Configure → Actions → Add action
2. Schema URL: https://your-app.up.railway.app/api/openapi.json
3. ChatGPT auto-discovers all REST endpoints
4. Paste the chatgpt.prompt value from GET /api/system-prompt as your GPT's System Prompt

---

6. MCP Tools

All 8 tools are available at POST /api/mcp (JSON-RPC, StreamableHTTP).

Tool	What it does	Who computes the data
get_state	Fetch full profile, plans, schedule, progress	—
save_state	Upsert any section of user state	—
log_completion	Log workout/diet, award XP + streak	—
normalize_user_input	Extract profile fields from freeform text	AI client passes extracted param
generate_plan	Save a diet or workout plan	AI client passes plan param
schedule_events	Save calendar events	AI client passes events array
get_history	Paginated completion history	—
export_report	Report in JSON / CSV / HTML	—

For normalize_user_input, generate_plan, and schedule_events: the tool description instructs the AI to generate the structured data itself and pass it as a parameter. The server validates and stores it. No server-side AI call is made.

Discover all tools and their input schemas:

POST /api/mcp
Content-Type: application/json
Accept: application/json, text/event-stream

{ "jsonrpc": "2.0", "id": 1, "method": "tools/list", "params": {} }

---

7. Security Notes

Current status: safe for personal use and demos.

• No authentication — any userId string can read/write any user's data. For personal use this is fine. For multi-user deployments, add an x-api-key header check before going public.
• Open CORS — app.use(cors()) allows all origins. Restrict to your domain in production.
• No rate limiting — add express-rate-limit if hosting publicly.
• SQL injection — safe, Drizzle ORM uses parameterized queries throughout.
• Secrets — no hardcoded credentials; all config via environment variables.

---

8. Environment Variables

Variable	Required	Default	Description
PORT	No	8080	Server port
DB_PATH	No	./data/fitness.db	SQLite file path — set to /data/fitness.db on Railway (volume mount)
PUBLIC_URL	No	Auto-detected	Base URL for report links — Railway sets this via RAILWAY_PUBLIC_DOMAIN automatically
LOG_LEVEL	No	info	Pino log level