AWS introduces Pizza Bot, an open source inbox for background AI agents

Amazon Web Services has released Pizza Bot, an open-source, self-hosted application designed to manage asynchronous AI agents. Originally tested by over 2,000 Amazon employees, the tool organizes background tasks into an email-style inbox with categories for history, completed work, and pending approvals. Built on LangGraph and SQLite, Pizza Bot supports major model providers like Anthropic and OpenAI, and integrates external tools via Model Context Protocol (MCP) servers with configurable human-in-the-loop controls.

Pizza Bot open source release

• ▪Amazon Web Services introduced Pizza Bot, a self-hosted, open-source application designed to manage background AI tasks that continue running while users work elsewhere.
• ▪Pizza Bot was originally used internally by more than 2,000 people at Amazon for tasks like meeting preparation, email drafting, and Slack summaries before being rebuilt as an open-source project.
• ▪Pizza Bot is licensed under the Apache 2.0 license and offers desktop builds for macOS, Windows, and Linux, alongside browser and terminal clients.

Asynchronous agent inbox architecture

• ▪Pizza Bot organizes background AI tasks into an email-style inbox with three primary categories: All for thread history, Unread for completed work, and Action for work paused for approval.
• ▪Tasks in Pizza Bot can be initiated manually, through cron schedules, or via webhooks, with the server managing scheduling and recording trigger occurrences durably.

LangGraph stateful execution runtime

• ▪Pizza Bot utilizes DeepAgents and LangGraph for stateful execution, using LangGraph checkpoints to retain thread state and approval pauses.
• ▪The Pizza Bot runtime relies on a Hono API server for execution and storage, using SQLite databases to hold cross-thread memory and application metadata.
• ▪Quitting the Pizza Bot desktop application stops its embedded server and ends active runs, whereas an always-on backend allows tasks to continue after the client disconnects.

MCP server tool integration

• ▪Pizza Bot integrates external tools using Model Context Protocol (MCP) servers, supporting existing Claude Code-compatible configurations.
• ▪In Pizza Bot, individual SKILL.md files define a worker's instructions and scoped tool access, making a skill callable only when its declared dependencies are available.

Configurable approval controls

• ▪Skill authors in Pizza Bot can configure specific policies using parameters like interruptOn and allowedDecisions to require manual user approval for specific tools.
• ▪Depending on configured tool policies, Pizza Bot users can choose to approve a proposed action, edit its arguments, or reject the action entirely.

Multi-provider model support

• ▪Users must configure their chosen model provider under Pizza Bot's settings before running background tasks.
• ▪Pizza Bot supports multiple AI model providers, including Amazon Bedrock, Anthropic, Google Gemini, OpenAI, OpenRouter, and local models via Ollama.