Amazon-developed Pizza Bot open-sources an inbox for long-running AI agents

Pizza Bot, an open-source inbox for background AI agents, has been released under the Apache 2.0 license. According to the project's GitHub repository, featured in a Show HN post on Hacker News, the application was developed at Amazon and refined through feedback from more than 2,000 internal Amazon users before being opened to the wider community.

An inbox with two queues

The core idea is to treat asynchronous agent work like email. You start or schedule a task, step away, and return to two views: finished runs collect in Unread, while runs parked awaiting a decision sit in Action. Agents keep working when you switch conversations or disconnect entirely, provided the api-server process stays up.

Because runs are checkpointed, they survive client disconnects, and cron or webhook triggers can start work without an open conversation. Threads can be grouped into folders without pulling their items out of the global Unread and Action queues. Human-in-the-loop approvals, long-term memory, file attachments, and desktop notifications are built into the workflow, with the stated goal of keeping people in control of consequential actions.

Runtime and clients

Pizza Bot runs on a stateful DeepAgents/LangGraph runtime and presents the same React interface in an Electron desktop app and in the browser. Desktop, web, and a terminal CLI all communicate with a shared api-server over HTTP and Server-Sent Events, and the project documents a standalone backend mode covering remote Electron clients, static browser deployments, Docker, Compose, and Kubernetes.

Skills, the extension mechanism, become tool-scoped subagents whose progress appears in an Activity panel, letting a task be delegated to a specialist agent. MCP servers and plugins can be added from the UI or configuration files, and a custom skill can override a built-in or plugin skill sharing its id without modifying the original.

On the code side, the repository splits the Hono api-server, CLI, Electron shell, and React web app into separate applications, with the LangGraph engine isolated in its own package; frontends consume protocol projections rather than importing runtime or model bindings directly.

Providers and configuration

A running api-server needs at least one model provider, though HTTP clients do not. Amazon Bedrock, Anthropic, Google Gemini, OpenAI, OpenRouter, and Ollama are supported, with models selected through a PIZZA_MODEL environment variable. Bedrock accepts an AWS profile, access keys, or a Bedrock API key, defaults to us-west-2 unless a region override or AWS_REGION is set, and routes models through Converse, OpenAI Responses or Chat Completions, or Anthropic Messages depending on the API family each model advertises. OpenAI and Anthropic also accept custom base URLs for compatible endpoints.

Installers ship for macOS on Intel and Apple silicon, Windows, and Linux on x64 and arm64, with SHA256 checksums attached to every release. The macOS builds are signed and notarized; the Linux packages are not signed, so the project advises checking them against the published sums. Running from source requires Node.js 24 or newer.

Security and local data

The design is local-first. The api-server binds to 127.0.0.1 by default, and any non-loopback binding requires authentication plus an explicit origin allowlist. Threads, checkpoints, memories, attachments, and logs live under a local data root, ~/.pizza-bot-oss by default, and model or tool requests go only to the endpoints you configure. Folder access is opt-in: individual read-only or writable folders are granted in settings, and the application receives no default access to the home directory.

The repository cautions that MCP servers and plugins are trusted code whose commands can execute with your user account's permissions, so they should be installed only from sources you trust. Secrets entered in the desktop app are protected with Electron safeStorage, and server-side configuration persists only environment-variable references rather than raw values.

Why it matters

As agents take on longer, multi-step tasks, the dominant chat interface — one conversation, one blocking request — fits the work less and less. Pizza Bot is a concrete attempt at a different pattern: treat agents as correspondents whose output and questions arrive asynchronously, then batch your attention into unread items and pending decisions. The Action queue in particular addresses a real gap by giving human-in-the-loop approvals a durable home rather than a dialog that disappears with the tab. With Amazon behind it, a permissive license, support for six model providers, and a documented path from local desktop to containerized backend, it doubles as a reference implementation that developers can inspect, fork, or adopt for their own background-agent workflows.