nutramap/foodPanelAI (name undecided)

Nutramap is a proof of concept for using LLMs heavily in data pipelines but in a transparent, auditable way. It's also the first ever nutrition tracker with an agentic interface. It can turn any chatbot into the best nutritionist money can buy, guaranteed. The web app is hosted live on nutramap.vercel.app but to test agent capablities you'll have to run it locally. If you'd like to see it in production please add your email to the list.

Agent Setup

Use this if you want an assistant to chat about meals, ask follow-up questions, and log entries using Nutramap. First, clone this repo to your desktop. If you're using Codex or Claude Code, the binary is recommended. If you're using an LLM without shell permissions (like bog-standard ChatGPT) you'll have to run the MCP server.

Binary quickstart

1. Install binary

chmod +x install-binary.sh
./install-binary.sh

2. Verify from terminal

Open a new terminal and test out the CLI.

foodpanel auth login
foodPanel log "one small cup vegan chilli"
foodpanel today

MCP server quickstart

chmod +x run-mcp.sh
./run-mcp.sh

Claude Code with CLI

Make sure you have the binary installed.

1. Install the local skill file for Claude Code:

mkdir -p ~/.claude/skills/foodpanel-cli

Now, either copy the skills file over to the claude code's folder...

cp /path/to/nutramap/skills/foodpanel-cli/SKILL.md ~/.claude/skills/foodpanel-cli/SKILL.md

Or set up symlink so updates to the repo are picked up automatically in every new chat (recommended if you keep the repo at a stable path):

ln -sf /Users/divyavenn/Documents/GitHub/nutramap/skills/foodpanel-cli/SKILL.md ~/.claude/skills/foodpanel-cli/SKILL.md

2. (Optional but highly recommended) Pre-approve foodpanel commands so Claude Code never prompts for permission:

Create or edit .claude/settings.json:

{
  "permissions": {
    "allow": [
      "Bash(foodpanel *)",
      "Bash(foodPanel *)"
    ]
  }
}

You can also just start claude code and tell claude to do this step.

3. Restart Claude and invoke the skill:

Use $foodpanel-cli for this session. Check what we've eaten today.

Codex with CLI

Make sure you have the binary installed.

These steps assume Codex Desktop UI (no manual config file edits).

1. From sidebar, open Skills.
2. Use skill installer and tell it to install the skill from https://github.com/divyavenn/nutramap/tree/main/skills/foodpanel-cli
3. Select this folder:
   • /path/to/nutramap/skills/foodpanel-cli
4. Confirm installation and enable the skill.
5. Restart Codex Desktop.

Then start a chat with:

Use $foodpanel-cli for this session.

ChatGPT Desktop

Make sure you have the MCP installed and running.

• Open ChatGPT Desktop.
• Go to Settings.
• Open Connectors / Tools / MCP Servers.
• Click Add server.

Configure:

• Name: foodpanel
• Command: uv
• Arguments: run --project /path/to/nutramap foodpanel-mcp (Use your own local repo path for /path/to/nutramap.)
• Environment variable:
  • FOODPANEL_API_URL=https://divyavenkatraman234--nutramap-backend-serve.modal.run

Restart ChatGPT Desktop.

Open a dedicated chat for this. Put this in the system prompt.

Use $foodpanel-agent for this session.

Check if it works by saying:

Call session_info now and show the tool result.

Claude Code with MCP

Make sure you have the MCP installed and running.

These steps are for the Claude Code CLI workflow (not desktop UI).

1. Register local Foodpanel MCP (one-time):

cd /path/to/nutramap
claude mcp add foodpanel -- uv run --project /path/to/nutramap foodpanel-mcp

cd /Users/divyavenn/Documents/GitHub/nutramap claude mcp add foodpanel -- uv run --project /Users/divyavenn/Documents/GitHub/nutramap foodpanel-mcp

2. Verify configuration:

claude mcp list
claude mcp get foodpanel

3. Install the local skill file for Claude Code:

mkdir -p ~/.claude/skills/foodpanel-agent
cp /path/to/nutramap/skills/foodpanel-agent/SKILL.md ~/.claude/skills/foodpanel-agent/SKILL.md

Or symlink so updates to the repo are picked up automatically in every new chat (recommended if you keep the repo at a stable path):

mkdir -p ~/.claude/skills/foodpanel-agent
ln -sf /path/to/nutramap/skills/foodpanel-agent/SKILL.md ~/.claude/skills/foodpanel-agent/SKILL.md

4. Start Claude Code in your project and invoke the skill:

Use $foodpanel-agent for this session. Call session_info now and show the tool result.

Codex with MCP

Make sure you have the MCP installed and running.

These steps assume Codex Desktop UI (no manual config file edits).

1. Open Codex Desktop.

2. Go to Settings.

3. Open MCP Servers.

4. Click Add Server and configure:

   • Name: foodpanel
   • Command: uv
   • Args: run --project /path/to/nutramap foodpanel-mcp

5. Save and enable the server.

6. From sidebar, open Skills.

7. Use skill installer and tell it to install the skill from https://github.com/divyavenn/nutramap/tree/main/skills/foodpanel-agent

8. Select this folder:

   • /path/to/nutramap/skills/foodpanel-agent

9. Confirm installation and enable the skill.

10. Restart Codex Desktop.

Then start a chat with:

Use $foodpanel-agent. I had one small cup vegan chilli. Ask follow-up questions if needed, then log it.

Verify it's running

In chat, run a simple tool call request like:

Use $foodpanel-agent for this session. Call session_info now and show the tool result.

If it returns server info (base_url, has_access_token), MCP is working.

Deploy MCP to Modal (Remote, Stateless)

This repo now includes a dedicated Modal app for MCP at:

• foodpanel_mcp/modal_app.py

Deploy it:

pip install modal
modal setup

modal deploy foodpanel_mcp/modal_app.py

If you need a different backend URL, set FOODPANEL_API_URL in Modal app environment settings.

After deploy, your MCP endpoint is:

• https://<your-workspace>--foodpanel-mcp-serve.modal.run/mcp

Auth behavior:

• The MCP server is stateless in HTTP mode.
• It does not persist user sessions to disk.
• Send user auth on each request with:
  • Authorization: Bearer <foodpanel_access_token>, or
  • x-foodpanel-access-token: <foodpanel_access_token>

In stateless mode, login returns an access_token so clients can cache and reuse it.

Developer Setup

🚀 Quick Start

Local Development (Recommended)

# Run both frontend and backend locally
docker compose -f compose.dev.yaml up -d

# Frontend: http://localhost:3001
# Backend: http://localhost:8001/docs

Production Deployment

# Deploy backend to Modal (serverless)
cd backend
./deploy-modal.sh

# Deploy frontend with Docker
docker compose up -d

📖 See DEPLOYMENT.md for complete deployment guide

---

🛠️ Manual Setup (Alternative)

Backend Package Management (using UV)

cd backend
uv sync              # Install dependencies
uv run uvicorn main:app --reload  # Run backend

Frontend Setup

cd frontend
npm install          # Install dependencies
npm run dev          # Run frontend

Manual Inspection of Databases

• MongoDB: Use MongoDB Compass or Atlas web interface
• Backend API: http://localhost:8001/docs (FastAPI Swagger UI)