Pizza Bot

• Starting a research or writing task, closing the window, and reviewing the result in Unread later
• Scheduling a daily briefing or report as a cron automation
• Kicking off an agent run from another system through a webhook trigger
• Letting an agent work through files in a granted project folder, with a skill that requires approval before it writes
• Running the backend on an always-on server and connecting from the desktop app, browser, or CLI

Pizza Bot is built for work that takes long enough that you do not want to watch it. The inbox, with finished work in Unread and approvals in Action, plus cron and webhook triggers, makes background agents easier to manage than a stack of chat tabs. Its SECURITY.md also spells out its data and network boundaries in detail. It is still a young community project, though, with no AWS support and a thin extension ecosystem, and always-on use means running and securing your own backend. It suits technical users comfortable with that more than teams looking for a supported product.

About Pizza Bot

Pizza Bot changes where you wait for an agent. Instead of watching a chat window, you start a task and move on. Finished work lands in Unread, and tasks that need a decision or approval wait in Action. All keeps the full history, and folders organize threads without hiding them from those queues. Tasks can be started by hand, on a cron schedule from the Automations screen, or by a webhook protected by its own secret. The runtime is built on DeepAgents and LangGraph, with checkpoints in SQLite, so a run survives a client disconnect. The agent can list, read, write, edit, and search files in its own scratch space and in folders you grant, run code in a sandboxed JavaScript interpreter with no network or filesystem access, drive a browser through the bundled Playwright MCP plugin, and keep optional long-term memory. Skills (SKILL.md files) become specialist subagents with scoped tool access, and you can follow each one's transcript and tool calls in the Activity panel. A skill can require approval for specific tools, and the approval card lets you approve, edit the arguments, or reject. You bring your own model: Amazon Bedrock, Anthropic, Google Gemini, OpenAI, OpenRouter, or Ollama, with custom base URLs for OpenAI- and Anthropic-compatible endpoints. It ships as an Electron desktop app for macOS, Windows, and Linux, and the same backend serves a browser client, a terminal CLI, and standalone deployments with Docker, Compose, or Kubernetes. Data stays in a local folder (~/.pizza-bot-oss by default), the server binds to localhost unless you configure a token and allowed origins, and the project includes no telemetry. The tradeoffs are maturity and scope. AWS released it on September 10, 2026 as a community project with no AWS support or SLA, and its authors say its skill and MCP ecosystem is still young. Agents run only while the api-server process is running, and quitting the desktop app stops the server it started, so always-on work needs a machine that stays up. MCP servers and plugins run as trusted code with your user's permissions.