What it does

PantryPilot Strands turns household food planning into one autonomous workflow. A Strands Agents SDK agent connects to PantryPilot MCP to remember inventory and preferences, plan meals, calculate shopping gaps, find products, prepare a reversible mock cart, and recall state. It does routine work quietly and surfaces decisions only for intent, allergy, budget or irreversible-action boundaries.

How we built it

The new agent layer is TypeScript on Node 22 using @strands-agents/sdk 1.17.0. Agent runs the model loop; McpClient connects to the Streamable HTTP MCP endpoint as the tool provider. toolExecutor: 'sequential' preserves ordering. The backend provides pantry, meal planning, shopping, products, cart draft and kitchen_run over durable SQLite state. Amazon Bedrock is the default model path. Credential-free judge-check discovers tools and invokes session_recall via Strands.

Why it matters and safety

This is not a recipe chatbot; it handles remember, plan, find gaps, prepare and preserve-context end to end. Allergy and budget gates are enforced. The cart is reversible; no real order, card charge or funds transfer occurs.

Pre-existing work disclosure

The PantryPilot MCP/domain server was created on 2026-09-10 and 2026-09-11 for a separate Amazon developer hackathon. For Agents for Humans, we created the new Strands runtime, native MCP integration, agent policy, judge checks, architecture and submission package. See agents-for-humans/PREEXISTING.md.

Challenges and accomplishments

We proved Strands—not a hard-coded workflow—owns reasoning while MCP owns deterministic state/actions. Highlights: real Strands agent, native MCP integration, stateful context, kitchen_run, safety boundaries and credential-free verification.

Demo video disclosure

The fallback video shows the pre-existing PantryPilot flow and predates the Agents for Humans Strands layer. It is not a Strands Agents SDK demo.

License

MIT.

Built With

Share this project:

Updates

Submission history