# Invarail

Invarail is a local, self-hosted authority layer for AI agents that separates what an agent is allowed to do from what it can learn to do, sitting in front of interchangeable workers for coding, repeatable procedures and open-ended tasks.

Invarail runs entirely on the user's own hardware and is deliberately layered: an authority plane holds permissions, target-bound grants, a confirmation ledger and audit trails that the agent cannot modify from inside, so learning may inform what it does but can never expand what it is allowed to do. Above that sits a daily-driver layer offering chat with graph memory, research reports, briefings, scheduling and image generation through adapters for Discord, Telegram, Gmail, the web and Chrome.

Execution itself is split between two patterns used side by side: known procedures run through a router and deterministic pipelines so the model only extracts parameters, while open-ended problems are handed to a small model with a bounded set of general tools such as read, edit, write and a scoped shell, inside acceptance criteria it must satisfy.

It targets small local models in the 7 to 30 billion parameter range and is written in TypeScript, released under the MIT license.

Website: https://github.com/PeterGreenAppliedAI/Invarail