AWS introduced Pizza Bot, as a self-hosted application for AI tasks that continue while users work elsewhere. It organizes completed results and pending decisions into an email-style inbox. Earlier versions served more than 2,000 people inside Amazon, supporting meeting preparation, email drafting, Slack summaries, CRM logging, and research. The public application was rebuilt as an open source project.

Deployable: Yes. Pizza Bot offers macOS, Windows, and Linux desktop builds, and browser and terminal clients connected to a local or standalone backend. Its code is licensed under Apache 2.0.

An Inbox for Asynchronous Work

Pizza Bot separates tasks into All, the thread history; Unread, completed work awaiting review; and Action, work paused for approval or an answer. Users can organize threads into folders and inspect delegated workers in the Activity panel. Tasks can start manually, through cron schedules, or through webhooks.

The server owns scheduling. After downtime, missed cron intervals produce 1 catch-up run instead of replaying every missed interval. Trigger occurrences are recorded durably.

How the Runtime Works

The application uses DeepAgents and LangGraph for stateful execution. A Hono API server owns runtime execution and storage. Electron and browser clients share a React interface, while all clients communicate with the server over HTTP and server-sent events. LangGraph checkpoints retain thread state and approval pauses; separate SQLite stores hold cross-thread memory and application metadata. Reconnecting clients can replay buffered events.

Closing a thread or disconnecting a client does not stop a running server. However, quitting the desktop app stops its embedded server and ends active runs. Checkpoints preserve the thread, but the step in flight can be lost. An always-on backend is required for work to continue after that desktop app exits.

Skills, Tools, and Approval Controls

Pizza Bot supports Amazon Bedrock, Anthropic, Google Gemini, OpenAI, OpenRouter, and Ollama. Configure a provider under Settings > Providers before running tasks.

The agent has scratch-file operations and a sandboxed JavaScript interpreter without network or host-filesystem access. It can delegate through task when ready skill workers exist. The filesystem layer separately supports explicit folder grants and persistent memory.

MCP servers expose external tools. Each SKILL.md defines a worker’s instructions and scoped tool access. A skill becomes callable only when its declared dependencies are available. Existing Claude Code-compatible .mcp.json configurations are supported, and plugins package skills with MCP servers.

Skill authors configure interruptOn and allowedDecisions to require approval for specific tools. Depending on that policy, users can approve, edit proposed arguments, or reject an action. These controls must be configured for the relevant tools.

Interactive Explainer

Run the illustrative custom-skill workflow below. Compare an always-on backend with an embedded desktop server, close the client during execution, and approve, edit, or reject the proposed action. Animation timing is illustrative; no external actions occur.

PIZZA BOT / BACKGROUND WORK

Illustrative custom skill

Client:

Open

Server:

Running

Checkpoint:

None

Worker

Approval

Result

Start a research brief, then close the desktop while it runs.

Prepare a research brief

The custom skill gates its publish tool with interruptOn.

Simulation only. No external calls.

Approval specification

Key Takeaways

1. An inbox for background agents: Pizza Bot organizes task history in All, completed work in Unread, and requests for approval or input in Action. Tasks support manual, cron, and webhook triggers.
2. Persistent execution with DeepAgents and LangGraph: Checkpoints preserve thread state and approval pauses. Tasks continue after client disconnection while the backend remains running.
3. Multiple model providers: Pizza Bot supports Amazon Bedrock, Anthropic, Google Gemini, OpenAI, OpenRouter, and local models through Ollama. Configured providers and tools can receive task data.
4. Scoped skills and configurable approvals: MCP servers expose tools, while SKILL.md files define specialist workers. Tool-specific policies let users approve, edit, or reject proposed actions.
5. Self-hosted and deployable: Apache 2.0 code, desktop builds, and standalone backend options are available. Each SQLite data directory supports 1 backend process.

---