🥘 PantryPilot — a privacy-first kitchen concierge

Track: Concierge Agents · Kaggle AI Agents: Intensive Vibe Coding Capstone

PantryPilot is a multi-agent personal assistant that answers three questions every household faces weekly:

1. What do we actually have at home?
2. What should we cook this week — given our allergies, diet and budget?
3. What exactly do we need to buy, and at which store?

The problem

Meal planning is a recurring, unpaid coordination job. Doing it badly costs money (duplicate purchases, food waste) and can be dangerous: dietary restrictions and allergies are health data, and forgetting them in a plan has real consequences. Generic chatbot planning fails on both fronts — it doesn't know your pantry, and pasting your health information into a cloud service every week is a privacy anti-pattern.

The solution

A team of specialized agents coordinated by an orchestrator, in which the LLM never owns your data. All personal state lives in a local SQLite database controlled by a custom MCP server ("PantryVault"). Agents can only interact with it through a narrow, validated, audited tool surface — and each agent only sees the tools it needs.

Architecture

Architecture

flowchart TB
    U["User"] <--> R
    subgraph ADK["Google ADK · Gemini"]
        R["pantry_concierge<br/><i>root orchestrator — no data tools</i>"]
        R -- delegate --> A["pantry_manager<br/><i>inventory CRUD</i>"]
        R -- delegate --> B["meal_planner<br/><i>weekly plans + preferences</i>"]
        R -- delegate --> C["shopping_assistant<br/><i>plan − pantry = list</i>"]
    end
    subgraph VAULT["PantryVault — custom MCP server (stdio, local)"]
        V["11 validated tools<br/>+ audit log"]
        DB[("SQLite<br/>pantry · preferences/allergies<br/>meal plans · shopping lists")]
        V --> DB
    end
    A -- "allowlist: 4 tools" --> V
    B -- "allowlist: 5 tools" --> V
    C -- "allowlist: 5 tools" --> V

Loading

Why agents? The workflow is genuinely multi-step and stateful: inventory changes daily, plans depend on preferences and inventory, shopping lists depend on plans minus inventory. Each specialist has a distinct instruction set and a distinct (minimal) tool permission set — a single prompt cannot enforce that separation; an agent architecture can.

Security (Concierge track focus)

Layer	Mechanism	Where
Data locality	All personal data in local SQLite; MCP over stdio — no ports, no network	pantry\_mcp/server.py
Least privilege	Per-agent MCP tool\_filter allowlists; root agent has zero data tools	pantry\_concierge/agent.py
Input validation	Type/length/bounds checks + closed key sets on every tool argument	pantry\_mcp/server.py
Injection safety	Parameterized SQL only; malformed JSON rejected before persistence	pantry\_mcp/server.py
PII redaction	before\_model\_callback strips emails/phones/IBANs from user input before any model call	pantry\_concierge/callbacks.py
Transparency	Audit log of every tool invocation (names + timestamps, never argument values), queryable by the user in chat	pantry\_mcp/server.py

Course concepts demonstrated

Concept	Where
Multi-agent system (ADK)	1 orchestrator + 3 specialists, sub\_agents delegation — code
MCP server	Custom-built PantryVault server (FastMCP, stdio) — code
Security features	Table above — code + video
Antigravity	Generated + ran the validation test suite (55 tests) — video
Deployability	adk web locally; Cloud Run instructions below — video

Setup

git clone <this-repo> \&\& cd pantrypilot
python -m venv .venv \&\& source .venv/bin/activate   # Windows: .venv\\Scripts\\activate
pip install -r requirements.txt

cp pantry\_concierge/.env.example pantry\_concierge/.env
# edit .env and set GOOGLE\_API\_KEY (free: https://aistudio.google.com/apikey)

adk web        # open http://localhost:8000 and pick "pantry\_concierge"

No other services required — the MCP server is spawned automatically as a subprocess, and the database file is created on first use at data/pantry.db.

Try this demo script

1. "I just bought 1kg couscous, 2 cans of chickpeas, 500g carrots and a bottle of olive oil"
2. "I'm allergic to peanuts, vegetarian on weekdays, budget 60 euros per week, I shop at Rewe and Colruyt"
3. "Plan my dinners for next week"        → uses pantry + respects allergies
4. "Save it and make my shopping list"    → plan minus pantry, grouped by store
5. "What did you do with my data?"        → agent reads back the audit log

Deployment (optional)

The agent runs on Cloud Run via the ADK CLI:

adk deploy cloud\_run --project <PROJECT\_ID> --region europe-west1 pantry\_concierge

For a private, single-user deployment keep --no-allow-unauthenticated (default) so the endpoint requires IAM auth; the SQLite vault should then be placed on a mounted volume (PANTRY\_DB=/mnt/data/pantry.db).

Repository layout

pantrypilot/
├── pantry\_concierge/        # ADK agent package
│   ├── agent.py             # orchestrator + 3 specialists, tool allowlists
│   ├── prompts.py           # all agent instructions in one reviewable place
│   ├── callbacks.py         # PII redaction before\_model\_callback
│   └── .env.example
├── pantry\_mcp/
│   └── server.py            # PantryVault MCP server (SQLite + validation + audit)
├── data/                    # created at runtime, gitignored
└── requirements.txt

Limitations & next steps

• Recipes come from the model's knowledge; a curated recipe MCP tool would make nutrition data exact.
• Store grouping is preference-based; live store inventory/price APIs are out of scope (and would leak shopping behavior — a deliberate trade-off for privacy).
• Multi-user households: per-member allergy profiles are a natural extension of the preferences schema.