Pizza Bot is an open-source inbox for background AI agents, published by Amazon developers
Pizza Bot is a self-hosted inbox that keeps AI agents running in the background and shows their work the way a mail client shows messages, published on 4 October by a team of Amazon developers under Apache 2.0 with 419 stars at launch.
Image: GitHub
Why it mattersAn agent that is only useful while you watch it blocks the human for the entire run, so an inbox pattern that handles scheduled runs, approval gates and multiple model providers is a template a team can copy for internal tools instead of writing one.
An agent that is only useful while you watch it does not do anything a chat window did not already do. Pizza Bot is an open-source project, published on 4 October by Amazon developers, that keeps agents running in the background and shows their work the way a mail client shows messages.
The repository at github.com/pizza-bot-app/pizza-bot describes it as "a local-first inbox for long-running AI agents, built with DeepAgents and LangGraph." It carries an Apache 2.0 licence and 419 stars at the time of this writing. InfoQ reports that principal technologist Joseph Dolivo started it as a side project called JoeBot in April 2025, that Igor Fil, Flávio Schuindt, Jacob Wert, Michael Karachewski and Itzik Paz worked on it with him, and that over 2,000 Amazon employees tested it internally before the release. The project is open source from Amazon developers and is not an official AWS service.
What is in the box
Three interfaces connect to one API server: an Electron desktop app with signed and notarised macOS builds, a browser client, and a terminal CLI meant for scripts and remote backends. Items appear in three queues, All, Unread and Action, where Unread is a completed run and Action is a run that has stopped and is asking for a decision. Checkpoints persist, so a client disconnect does not stop the work, and the run picks up on the next connection.
Model choice is set with an environment variable, PIZZA_MODEL=<provider>:<id>, and the supported providers are Amazon Bedrock, Anthropic, Google Gemini, OpenAI, OpenRouter, Requesty, and Ollama. MCP servers and Claude-style Agent Skills plug in as specialised sub-workers that the main agent can delegate to. Files and state are kept on the local machine, and folder access is granted per folder rather than through one broad permission.
Why this shape is interesting
The common pattern for an agent today is a chat window the human watches. That blocks the human for the whole run, pushes long tasks off the table because nobody will sit through them, and makes approval gates feel like interruptions rather than work. An inbox flips it: the human reads completed runs in batches, and the agent asks for a decision only when it needs one. The feature list on the Pizza Bot repository, "schedules, webhooks, approvals, long-term memory", is what that pattern needs to work.
The reason to read this repository is not the stack. It is the file layout, the queues and the approval protocol. A team that wants to run its own agents against its own data has a working template now, under a permissive licence, from Amazon developers who already shipped it to over 2,000 internal users.
Source
Primary source: pizza-bot-app/pizza-bot on GitHub. Reporting: InfoQ, "Pizza Bot: Open-Source Inbox for Background AI Agents".
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
