Email-style inbox for agents comes out of Amazon

Pizza Bot, an AWS project that went open source on September 10, puts finished agent work in your inbox as unread threads and moves anything that needs a person into a separate queue called Action. The agents keep working whether or not anyone has the window open, The New Stack reported. The creators sum up the premise in one line: “The interface assumes you are not watching.”

That line comes from the AWS launch post, which argues that live chat is the wrong fit for this kind of work. Chat assumes both parties are present. That falls apart when a task takes several minutes, waits on an approval, or runs on a schedule while the user is away. So Pizza Bot is shaped like an email client, and a thread is a piece of work you come back to rather than a session you have to sit through.

From an MCP server to an inbox

In April 2025 Joseph Dolivo, principal technologist at AWS Startups, started what he calls a “side-of-desk passion project” named JoeBot. It automated repetitive CRM logging. He then teamed up with his colleague Igor Fil, and they turned the script into Pizza Bot: an MCP server that ran parameterized, deterministic “recipes” across Amazon’s internal systems.

People outside the team that built it started using it. Dolivo says it eventually grew to more than 30 contributors and over 2,000 users inside Amazon. AWS says those internal users relied on it for meeting preparation and follow-ups, email drafting, Slack summaries, CRM logging, planning the day, and web research.

“An MCP server requires an MCP client, and expecting non-technical users to work out of an IDE or terminal was never going to cut it,” Dolivo says. “We had to meet people where they actually work and own the experience end to end.”

The result is a desktop app for macOS, Windows and Linux, built around an inbox, with browser and terminal clients alongside it. AWS says the team rebuilt it from the ground up as open source, and that coding agents made the rebuild practical for a small team. The project has also left Amazon. It lives in its own GitHub organization as a standalone community project and comes with no AWS support or service-level agreement. You host it yourself.

Unread, Action, and the Activity panel

According to the team, four requirements came up in every workflow they built. An agent had to check several sources, prepare a result in the background, ask before taking a consequential action, and come find the person who asked once it was done. The inbox is built around the last two.

Work can start three ways: by hand, on a schedule, or from a webhook. Results land in three queues. All holds the thread history. Unread holds finished work nobody has looked at yet. Action holds runs paused while they wait for an approval or an answer. The Activity panel shows jobs handed off to specialist agents, including their tool calls, progress and their own transcripts. Each skill becomes a subagent limited to its own tools, the project README explains, and its progress shows up in that panel.

“Instead of you having to initiate every conversation or wait on a prompt, scheduled agents run autonomously in the background and surface updates directly into your inbox for review and triage,” Dolivo wrote in a LinkedIn post on Thursday.

The creators go further in the launch post. “Nothing else we’ve seen starts there,” they write, “and that one assumption is what buys you pauses that outlast the session that created them, notifications worth acting on, and scheduled work that produces threads instead of logs.” LangChain was working on the same idea back in January 2025, when its CEO, Harrison Chase, coined the term “ambient agents” for agents that react to events, run concurrently and bring in a human only when needed. LangChain’s reference implementation was an email assistant built on LangGraph, and the company also built an “Agent Inbox”: a standalone interface, modelled on email and customer-support software, for tracking open interactions between people and background agents. That makes the “nothing else” claim a hard one to defend.

What keeps a paused run alive

The overlap goes down into the code. Pizza Bot’s agent runtime is built on DeepAgents, LangChain’s open-source harness for long-running agent tasks, which in turn runs on LangGraph, LangChain’s runtime for stateful agents. Persistence is what matters here. LangGraph saves a checkpoint of the agent’s state as the run goes, so a run can stop for an approval, outlive a client that disconnected, and pick up later without starting over. Pizza Bot keeps those checkpoints, along with threads and other app data, on the local machine in SQLite and ordinary files.

AWS has its own open-source agents SDK, Strands Agents, available since May 2025. The New Stack points to an AWS sample repository that calls Strands a “lighter-weight alternative to LangGraph for agents that don’t need explicit graph control flow.” Pizza Bot leans heavily on exactly that graph machinery: LangGraph’s persistence and interrupts drive its checkpoints, pauses and resumable threads. The outlet also offers a simpler explanation. Since Pizza Bot is a community project and not an AWS service, there may never have been any expectation that it use Strands.

You pick the model behind it: Anthropic, Amazon Bedrock, Google Gemini, OpenAI, OpenRouter, or a local model through Ollama. You extend the app with MCP servers and Agent Skills. The bundled browser-automation skill, for example, uses Playwright MCP to navigate and interact with websites. Human-in-the-loop approvals, long-term memory, file attachments and desktop notifications are built in, the README says.

By default the desktop app starts a local api-server, and per the README that process is the only thing that has to stay up. The same React interface runs in Electron and in the browser, and the desktop, web and terminal clients all talk to the api-server over HTTP and server-sent events. Closing a window or losing the connection doesn’t stop the agents. Move the server to an always-on host or into a container, and scheduled agents keep working with the laptop shut, their threads waiting to be opened later from another device.