Rat Things is a self-hosted control plane for AI agents. It runs per-session EC2 or Lambda MicroVM workers in your own AWS account, plus a local workspace on your computer. It is built for developers making coding agents and services that need saved work, isolated execution, and control over their infrastructure. It is released under the MIT license.
Rat Things saves conversations and work so you can come back later, continue after a worker stops, or run tasks on a schedule. You can drive it from a native Rust desktop app, a CLI, or an interface compatible with the OpenAI Agents API, so the official OpenAI SDK can talk to a deployment you own.
You can start at home. Local Sessions run on your computer with private storage, using Codex or Pi Durable with your ChatGPT subscription. When you are ready for the cloud, a single quickstart command checks your AWS identity and model access, asks before deploying, and sets up Bedrock, private AWS endpoints, and no public runtime egress by default. An optional AWS spending alert can be added at install time.
Two runtimes are supported. Codex is the default, and Pi Durable can be selected per Session. Pi Durable saves conversations, pending tasks, and tool results, and a replacement worker can recover committed work without depending on the original machine.
Other features described in the project README:
On security, the agent process runs as a separate non-root user inside a Lambda MicroVM or dedicated EC2 worker, worker ingress is private, prompts and results live in encrypted S3, and integration credentials stay behind host-side brokers. Capabilities are fixed before launch.
Rat Things is an engineering preview. The repository documents current capabilities and limits, and includes comparison guides against E2B, Daytona, Modal, and Bedrock AgentCore.
Repo (MIT): https://github.com/gpazo/Rat-Things