Originally published at https://tekmag.thsite.top/aws-pizza-bot-open-source-inbox-for-background-ai-agents/
Pizza Bot is a self-hosted, Apache 2.0 open-source application from AWS engineers that runs long-running AI agents in the background and hands their results to you through an email-style inbox. You assign work, close the laptop, and come back to finished threads plus a queue of items that need your approval, rather than watching a chat window the whole time.
I verified the details directly against AWS's Open Source blog post announcing the release and cross-referenced with the GitHub repository. The numbers hold up: v1.1.0 shipped on September 19, 2026, the repo sits at 327 stars and 25 forks, the license is Apache 2.0, and every thread and file lands in one local folder at ~/.pizza-bot-oss.
Key Takeaways
- Pizza Bot is a finished application, not a framework: an inbox with three queues (All, Unread, Action) plus desktop, browser, and terminal clients for agents that keep working after you disconnect.
- It is built on DeepAgents and LangGraph rather than AWS's Strands SDK, stores everything locally in SQLite and plain files, and supports Anthropic, Bedrock, Gemini, OpenAI, OpenRouter, and Ollama.
- The project started inside Amazon in April 2025 as "JoeBot," was used by 2,000+ employees, and went open source with Amazon-specific integrations deliberately dropped for the community to fill.
- It ships as a community project with no SLA: the team calls itself "Chefs," after Amazon's two-pizza teams.
What Pizza Bot Is
Pizza Bot is a self-hosted, open-source inbox for AI agents that work in the background. AWS's Open Source blog framed the launch that way: agents run threads over hours or days, and the interface you actually use is a queue of their output instead of a live chat session. Three queues organize that output. All holds the full thread history. Unread collects finished work you have not reviewed yet. Action parks items that paused while waiting on your approval or a decision, which is the only queue that demands you.
The project began inside Amazon in April 2025 as an internal side project called JoeBot, created with Joseph Dolivo to automate repetitive CRM logging. According to TechDebrief's account of the origin story, it grew to 30+ contributors and 2,000+ Amazon employees before being rebuilt as open source and moved into its own GitHub organization. The name comes from Amazon's two-pizza team model, and the development team calls itself "Chefs," led by Executive Chefs Joseph Dolivo and Igor Fil.
Two things separate Pizza Bot from another agent toolkit. First, the design assumption is that you are not watching: threads outlast sessions, an agent that finishes while your laptop is closed resumes from its checkpoint when you reconnect, and scheduled work produces threads you review later rather than log files you tail. Second, it ships as a complete application with Electron desktop clients for Windows, macOS, and Linux plus browser and terminal interfaces, not as a library you wire into your own UI. In other words, Pizza Bot turns agent tooling from something builders assemble into something knowledge workers can just open and use.
How Pizza Bot Is Built (Architecture)
Pizza Bot's architecture is a server-plus-client model running on DeepAgents and LangGraph, with all state checkpointed to SQLite and plain files inside one local folder (~/.pizza-bot-oss). The server can live on your own machine or on an always-on host or container, and the clients reach it locally or over the network. The New Stack's technical analysis of the release notes the stack deliberately skipped AWS's own Strands Agents SDK: LangGraph's mature tooling, broad ecosystem, and familiarity with Python developers reduce the friction of community contribution.
Checkpointing is the load-bearing part. Each thread saves its state, so a closed laptop does not kill the work; on reconnect the agent picks up where it left off. MarktechPost covered the release with emphasis on the local-first storage model: no database server to install, no cloud dependency for the data itself, just one folder you can move or back up. For anyone already building agent memory on top of a framework, that checkpoint-per-step pattern is close to the kind of persistent memory architecture we examined in our guide to building a memory operating system for AI agents, but pointed at a finished app instead of a research paper.
Model choice is yours. Pizza Bot supports Anthropic's Claude models, Amazon Bedrock, Google Gemini, OpenAI, OpenRouter, and Ollama for fully local inference. Provider credentials go to your operating system's secret store rather than plaintext config, and browser clients never receive them. For remote deployments you enable an API token plus an explicit allowed-origins list. The project ships with no telemetry at all.
What Pizza Bot Does (Features)
Pizza Bot extends through two mechanisms: Model Context Protocol (MCP) servers for tools, and Agent Skills using Anthropic's SKILL.md convention. Each specialist skill can define its own MCP tools and set per-tool approval policies with interruptOn, which is how work lands in the Action queue: a skill says "this tool needs a human before it runs," and the thread pauses until you approve. Claude Code-compatible .mcp.json files work as-is, so existing configurations port over. Two skills ship by default, a browser automation skill and a guide skill, and Amazon-specific integrations were dropped during the OSS rebuild because they only made sense inside the company.
Specialist agents can delegate work to one another, and the app shows the delegation in an Activity panel so you can see which subagent produced which result. For a practical look at how skill files themselves are structured in production systems, our breakdown of the major AI coding assistants and their skill conventions covers the same SKILL.md ground from the tooling side.
Automation rounds out the package. Cron-style schedules start threads on a timer, and secret-protected webhooks can trigger work from external systems. The missed-run rule is the one that trips people up. A schedule that fired while the machine was off runs exactly once on wake. It does not backfill a job for every missed occurrence, so a laptop that was closed for a week does not queue seven runs on open. When a specialist thread needs input, it stops in the Action queue instead of guessing. The approval model is what makes background agents safe to leave running: agents do the work, humans keep the veto.
How Pizza Bot Compares
Compared to hosted assistants like ChatGPT or Perplexity, Pizza Bot is self-hosted: your data stays in one local folder, there is no telemetry, and local models through Ollama mean no provider account is required at all. Compared to chat-based agent tools, the difference is the session model. A chat window assumes you are present; Pizza Bot assumes you are not, which is why threads checkpoint, why a closed laptop does not cancel work, and why approvals can arrive from the desktop app, browser, or terminal on any device. The harness-and-skill architecture here is an example of the wider pattern of treating agent harnesses as an engineering discipline, but packaged so no harness code needs to be written to get started.
The sharpest comparison is with AWS's own Strands Agents SDK, because the two are not competing products. Strands is a lighter SDK for developers embedding agents in applications they write themselves, while Pizza Bot is a finished product that needed explicit graph control flow and checkpointing, which is why it reached for LangGraph instead. The Hacker News discussion of the release turned on trust: people wanted agents to do real background work without a chat tab open, which is exactly the design bet the inbox makes.
Conclusion
Pizza Bot is what agent tooling looks like when the assumption is that you stop watching: background execution with checkpoints, three inbox queues that split finished work from work needing you, and human approval as a first-class step instead of an afterthought. The honest caveat is that it is a young community project, 327 stars and v1.1.0 as of September 2026, with no SLA behind it, so the Amazon pedigree is provenance, not a support contract.
If you found this analysis useful, share your take in the comments: are you running agents in the background yet, and what made you trust them? And if you build on top of frameworks rather than finished apps, start with our coverage of the AI coding assistant landscape.
Frequently asked questions
Is Pizza Bot an AWS-managed service?
No. It is an open-source community project released by AWS engineers under the Apache 2.0 license, with no SLA or AWS support. You host it yourself, and all data stays on your machine.
Can I use local models with Pizza Bot?
Yes. Ollama is a supported provider, so you can run the entire stack against local models with no external API account. Anthropic, Bedrock, Gemini, OpenAI, and OpenRouter are also supported.
What happens to a scheduled task if the machine was off when it was due?
A missed cron-style schedule fires exactly once when the machine wakes or the server starts. It does not replay one job per missed occurrence.
Why did Pizza Bot use LangGraph instead of AWS's own Strands Agents SDK?
Strands is a lighter SDK for embedding agents in applications you write yourself. Pizza Bot is a finished app that needed explicit graph control flow and checkpointing, where LangGraph's maturity, ecosystem, and developer familiarity make sense.
Where does Pizza Bot store its data?
Everything lives in one local folder, ~/.pizza-bot-oss, as SQLite databases plus plain files. Provider credentials are kept in the operating system's secret store, and the app collects no telemetry.
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": []
}
References
- Introducing Pizza Bot, an open source inbox for AI agents that work in the background (AWS Open Source Blog)
- pizza-bot-app/pizza-bot on GitHub
- AWS open-sources Pizza Bot: email-style inbox for background AI agents (The New Stack)
- Your AI agents' reports and questions have a new inbox, courtesy of AWS (The Register)
- Show HN discussion of Pizza Bot (Hacker News)
- AWS Introduces Pizza Bot, an Open-Source Inbox for Background AI Agents (MarktechPost)
- AWS Releases Pizza Bot, an Inbox for AI Agents That Work While You Sleep (TechDebrief)
Top comments (0)