Download the desktop build for macOS, Windows or Linux from the project's GitHub releases page, connect a model provider under Settings then Providers, and assign your first task — that is genuinely all it takes to start using AWS Pizza Bot, the open-source inbox for AI agents that AWS released on 10 September 2026. Instead of watching an agent grind through a task in a chat window or terminal, you hand work over, close the window, and the agent comes back to you when it has finished or when it is stuck — exactly like email from a very fast colleague.
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →
The details here come from the AWS Open Source Blog announcement of 10 September 2026, written by AWS's Joseph Dolivo and Igor Fil, plus the project's own GitHub repository. Pizza Bot is a community open-source project under the Apache 2.0 licence — free to use, self-hosted, and pointedly not an AWS paid service: there is no service-level agreement and no bill attached beyond whatever model provider you plug in.
What AWS Pizza Bot Actually Is
AWS describes Pizza Bot as "an open source application that runs AI agents in the background" with "an inbox, so the agent comes back to you when it's finished or when it's stuck". The mental model is deliberate: chat interfaces make you babysit the agent, while an inbox makes the agent report to you. Work is treated as a durable thread — an agent is assigned a task, runs independently, and surfaces only when there is something for you to see or decide.
The inbox itself is split into three queues:
- All — the full thread history of everything your agents have worked on.
- Unread — completed work you have not looked at yet.
- Action — work that is paused, waiting on your approval or your answer before the agent continues.
That third queue is the quietly important one. Long-running agents fail in practice not because models are weak but because somebody has to notice when a decision is needed — approve this purchase, confirm this email, choose between these two options. Putting those moments in a dedicated queue, rather than losing them in a scrolling chat, is the difference between an agent you trust with real work and a demo. It is the same human-in-the-loop pattern that managed platforms are converging on, as covered in the OpenAI Agents API public beta breakdown — except Pizza Bot does it self-hosted and free.
If you want background agents running your business rather than just your experiments, the AI Profit Boardroom gives you 1,000+ done-for-you agent workflows plus five live coaching calls a week — check out the AI Profit Boardroom. Want a 1-on-1 plan for automating your SEO and content first? Book a free SEO strategy session — it is free and you leave with next steps.
Models, Tools And Skills: What It Connects To
Pizza Bot is provider-agnostic. The AWS announcement lists support for Anthropic, Amazon Bedrock, Google Gemini, OpenAI, OpenRouter, "or even a local model through Ollama". That last option matters for cost control: a background agent that runs all day on a local model costs you electricity, not tokens. If you have not run a local model behind an agent before, the Hermes with Ollama setup guide covers the general pattern, and it transfers directly.
On the tools side, Pizza Bot connects agents to the outside world through Model Context Protocol servers — the same MCP standard the rest of the agent ecosystem has settled on, explained from scratch in the Hermes MCP server guide. And for packaging expertise, it supports Agent Skills, the Markdown convention Anthropic introduced, so a specialist workflow can be written once as a skill file and handed to any agent. According to the project's GitHub page, the whole thing is built on DeepAgents and LangGraph, and runs as a server plus a client — locally on your machine, on an always-on box, or in a container.
Where AWS Pizza Bot Came From
This is not a weekend hackathon project. AWS says earlier internal versions were used by more than 2,000 people inside Amazon for meeting preparation and follow-ups, email drafting, Slack summaries, CRM logging, day prioritisation and web research. That usage list doubles as a starter menu for what to delegate first: none of those tasks need frontier-model brilliance, all of them eat calendar time, and every one of them fits the fire-and-forget inbox pattern.
A sensible first week with Pizza Bot looks like this:
- Day 1 — install and connect one provider. Grab the build for your platform from the releases page, add one API key under Settings then Providers, and send a trivial task end to end so you see the Unread queue work.
- Day 2 — delegate one recurring chore. Web research briefs and meeting prep are the classic starters, because a late or imperfect result costs you nothing.
- Day 3 — add one MCP tool. Give the agent access to one real system you use, and route anything consequential through the Action queue rather than granting blanket permissions.
- Day 4 — write your first Agent Skill. Take a process you explain repeatedly, write it as a Markdown skill, and attach it.
- Day 5 — move it off your laptop. A container or an always-on machine turns it from a toy into infrastructure.
Structure is what separates agents that compound from agents that entertain, and that principle scales beyond one tool — it is the entire premise of the Agent OS, and the workflow patterns in the best Hermes workflows guide port over to Pizza Bot threads almost unchanged.
Is AWS Pizza Bot Free?
Yes — the application itself costs nothing. Pizza Bot is released under the Apache License 2.0, which permits commercial use, modification and redistribution, and AWS is explicit that it ships as a community open-source project without AWS support or a service-level agreement. What you pay for is whatever sits behind it: if you connect an Anthropic, OpenAI, Gemini, Bedrock or OpenRouter key, you pay that provider's normal token rates for the work your agents do, and nothing on top for the inbox. If you connect a local model through Ollama instead, the token bill disappears entirely and your only real cost is the machine it runs on.
That cost structure is worth pausing on, because background agents change how tokens get spent. A chat session ends when you close the tab; a background agent keeps working. Long tasks on an expensive frontier model can add up quietly, which is exactly why the provider-agnostic design matters — you can route all-day drudge work to a cheap or local model and reserve the premium key for tasks where output quality is the point.
Is AWS Pizza Bot Right For Your Stack?
Pizza Bot occupies a specific niche: self-hosted, single-operator-friendly, interface-first. Ranked by fit, the strongest use cases first:
- Solo operators and small teams who want background agents without a platform bill. This is the bullseye — Apache 2.0, your hardware, your keys, your data.
- Builders already invested in MCP and Agent Skills. Your existing tools and skill files plug straight in.
- Privacy-sensitive workflows. Self-hosting plus Ollama keeps client data entirely on machines you control.
- Teams needing scale, SLAs or managed orchestration. Weakest fit — that is what managed agent platforms are for, and a community project without support guarantees is the wrong foundation for mission-critical volume.
One honest caveat: an inbox app does not make the agents themselves smarter. Model choice still decides output quality — the Goldie Bench write-up covers how the current crop of model brains compare in hands-on tests, and it is worth reading before you decide which provider key to put behind your background workers. If you run multiple agents on one project, the coordination patterns in the Hermes agent swarm guide are the natural next step once a single inbox stops being enough.
The bigger signal is worth registering: AWS open-sourcing a local-first agent inbox — rather than shipping it as a paid service — confirms that background agents with human checkpoints are becoming default infrastructure, the way version control and CI once did. Getting comfortable with the pattern now, while it is free and simple, is how you avoid paying enterprise prices for it later. AWS Pizza Bot is one of the lowest-risk ways to start: install it, hand it one chore, and see whether an agent that reports to your inbox changes how much of your week you get back.
If you want the full library of agent workflows, prompts and daily tutorials that members use to put tools like this to work, check out the AI Profit Boardroom — join 3,000+ AI operators inside. Rather talk through your own automation plan 1-on-1? Book a free SEO strategy session and get a personalised route map.
