Dev tools

Pizza Bot, a new open-source app from AWS engineers, gives long-running AI agents an email-style inbox

September 10, 2026 at 5:20 PM PT

Pizza Bot desktop app showing an inbox of finished agent work and one approval request

Image: The New Stack

Why it mattersTeams building agents that take longer than a chat message can now start from a finished application with an interaction pattern people already understand, saving them from stitching one together on top of an SDK.

Pizza Bot, an Apache-2.0 desktop app for running long-running AI agents in the background, launched on 10 September at github.com/pizza-bot-app/pizza-bot. The project was built by Joseph Dolivo, principal technologist at AWS Startups, and colleague Igor Fil. The AWS launch post states plainly that Pizza Bot is a community project, lives in its own GitHub organisation, and carries no AWS support or service-level agreement.

The design starts from one sentence the launch post keeps repeating: "The interface assumes you are not watching." Finished work arrives in an Unread queue, anything the agent has paused to ask about arrives in an Action queue, and an Activity panel shows the specialist agents the main agent handed work to, with each specialist's own transcript. A chat window assumes both parties are present at the same time; Pizza Bot assumes the person who started the task has walked away.

How the run survives a closed lid

Pizza Bot is a server and a client. The desktop app for macOS, Windows and Linux wraps both, starting a local server on the same machine and connecting to it the way a remote client would. A run lives on the server, so closing the thread, reloading the browser client or switching to the terminal client leaves the run going, and reconnecting catches up on what was missed. Put the server on an always-on host or in a container, and the same threads are reachable from a laptop, a phone or another workstation.

The runtime underneath is DeepAgents on LangGraph, LangChain's open-source graph runtime for stateful agents. LangGraph checkpoints the run as it goes, so a pause waiting on approval, a disconnected client or a crashed model call keeps the thread and loses only the step in flight. State, checkpoints, memories, attachments and logs sit in one folder as SQLite databases and ordinary files; the launch post describes that folder as the one thing worth backing up.

Which model, which tools, which network

The model provider is the operator's choice: Anthropic, Amazon Bedrock, Google Gemini, OpenAI, OpenRouter or a local model through Ollama. Extensions come through MCP servers and through Agent Skills, the Markdown packaging format Anthropic introduced earlier this year; the bundled browser skill uses Playwright MCP. Provider keys go to the operating system's secret store, the launch post says, and are never shipped to a browser client.

Pizza Bot listens only on the local machine by default. A deployment that is not local-only requires an API token and an explicit list of allowed origins; the server refuses a remote client without both. The SECURITY.md in the repository documents the full boundary and the private disclosure path.

Where this came from

Dolivo tells The New Stack that the project started in April 2025 as a side script called "JoeBot" for CRM logging, was rebuilt with Fil into an MCP server that ran parameterized recipes across internal systems, and grew inside Amazon to more than 2,000 users and over 30 contributors before being rewritten from scratch as this week's open-source release. The AWS launch post confirms the internal user count and says the name is a reference to Amazon's two-pizza teams. Amazon already has an open-source agent SDK, Strands Agents, released May 2025; Pizza Bot's own repository frames Strands as the lighter-weight option for agents whose control flow does not need to be a graph, and the AWS launch post pitches Pizza Bot as a finished application for people who just want the work done. Anyone shipping an ambient-agent workflow now has a reference implementation to fork.

Source

Source: AWS Open Source Blog

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

More from AI News

BankMCP puts a read-only MCP server in front of Enable Banking, so Claude Code and Cursor can query your own bank accounts

A one-person MIT-licensed project called BankMCP wraps Enable Banking's PSD2 API in a self-hosted MCP server, so an MCP client such as Claude Code, Cursor, ChatGPT or Ollama can read balances and transactions from 2,700 European banks with no third party holding the data.

Source: GitHubDev tools

Engrim ships a local SQLite memory store that survives switching between Claude Code, Cursor, Antigravity and Codex

Engrim is a local-first SQLite memory engine for coding agents that carries architectural decisions and project state across sessions and across models, from Claude Code to Cursor to Google Antigravity to Codex CLI.

Source: Hacker NewsDev tools

AWS open sourced the agent workspace it built internally, after 39,000 Amazon developers picked it up without a mandate

Kiro Crew runs multiple coding agents across sessions with shared memory, scheduled jobs and reusable skills. Built inside Amazon as MeshClaw, it is now open source and runs locally or on your own infrastructure.

Source: PressDev tools