How to Self-Host Pizza Bot, AWS's Free Alternative to Paid Agent Inbox Tools
AWS ran Pizza Bot internally for 2,000+ employees before open-sourcing it. Here's how to install it, connect a model for free, and get your first background agent task running.
Contents9
AWS quietly used Pizza Bot inside the company for more than 2,000 employees before releasing it as a free, Apache 2.0 open-source project in September 2026. It's an inbox for AI agents: instead of watching a chat window, you assign work and it comes back to you the way email does, with an unread queue and an action queue for anything that needs your approval. Here's how to actually get it running, including the part most announcements skip: how to do it without paying for a single API call.
What you need before you start
You need a machine to run it on (macOS, Windows, or Linux all work) and a way to talk to a model. That second part is where the free path lives: Pizza Bot connects to Anthropic, Amazon Bedrock, Google Gemini, OpenAI, or OpenRouter if you want a hosted model, but it also connects to Ollama, so you can run a local model and never send a token to a paid API.
Step 1: Download and install
Grab the latest build for your operating system from the project's GitHub releases page. macOS builds are code-signed, so they'll open without extra prompts; Windows and Linux builds may ask you to confirm you trust the download, which is normal for a smaller open-source project without a code-signing budget.
Step 2: Connect a model provider
Open Settings and add credentials for whichever provider you want. If your goal is to run Pizza Bot for free, point it at Ollama and pick a local model that fits your machine's memory. You'll trade some capability for zero cost, which is a fair trade for routine tasks like drafting a Slack summary or logging a CRM update, and a worse trade if you want it handling anything that needs real reasoning.
Step 3: Run the two default skills first
Pizza Bot ships with exactly two skills out of the box: a browser-automation skill built on Playwright MCP, and a self-referential skill that answers questions about Pizza Bot itself. Run both before you write anything custom. They'll show you how the approval flow actually works in practice, which matters more than reading about it.
Step 4: Write your first custom skill
Once the defaults run out, add a `SKILL.md` file. Each skill defines what tools the agent can use, which of those tools need your approval before running, and whether you get an editable proposal instead of a yes/no prompt. Start narrow: a skill for one specific recurring task (weekly report drafting, meeting prep, whatever you actually do every week) teaches you the pattern faster than trying to build a general-purpose assistant on day one.
Step 5: Schedule it and connect webhooks
Pizza Bot supports cron-based scheduling and webhook triggers, so once a skill works when you run it manually, you can set it to run every morning or fire when an external event hits a webhook. This is where the "inbox" framing actually pays off: you stop opening a chat window entirely, and work just shows up as unread items when it's done.
Free vs. paid: what actually changes
| Local model (Ollama) | Hosted model (Anthropic/OpenAI/Gemini/Bedrock/OpenRouter) | |
|---|---|---|
| Cost | $0 beyond your own hardware | Pay-per-token, varies by provider and model |
| Capability | Limited by your machine and the local model's size | Full frontier-model capability |
| Data | Never leaves your machine | Sent to the provider you connect |
| Best for | Routine, low-stakes tasks (summaries, logging, drafts) | Anything needing strong reasoning or long context |
What Pizza Bot won't do for you
It's a community project, not an AWS product. There's no support ticket to file and no SLA if it breaks; AWS says as much in its own announcement. It also ships with fewer ready-made skills than the internal Amazon version had, so expect to write your own for anything beyond browser automation. If you want something with more established support out of the box, [Cline](https://questloops.com/tools/cline) and [Devin Desktop](https://questloops.com/tools/devin-desktop) are both more mature options, though neither is built around Pizza Bot's inbox-for-any-task model. Full details on pricing, features, and honest tradeoffs are in the [Pizza Bot review](https://questloops.com/tools/pizza-bot).
One more thing worth knowing before you hand an agent broad tool access: Anthropic's CEO recently [warned that AI agent swarms could pose a serious risk within 6-12 months](https://questloops.com/blog/ai-agent-swarm-warning-amodei-vs-security-skeptics), a claim security researchers have pushed back on, but the underlying point about approval gates stands regardless of whether the doomsday timeline is right. Pizza Bot's action queue is exactly the kind of pause-before-acting design that makes self-hosted agents safer to run unsupervised.
The takeaway
Pizza Bot is free to run end to end if you're willing to use a local model, and worth the setup time if you want an inbox-style way to delegate background work without a subscription. Just budget an evening for it: between picking a model, running the default skills, and writing your first real `SKILL.md`, it's more of a project than a five-minute install.