Octos is an embeddable AI agent harness kernel written in Rust: the runtime layer that agent applications are built on. Published on GitHub with about 760 stars under the Apache-2.0 license and updated in October 2026, it provides the execution loop, context management, memory, and tooling that agents need.
Rather than being another agent app, Octos is a reusable kernel you compile into your own application. It owns agent execution and runtime state — the loop, context management, memory, tools, skills, workflows, and coordination between agents — while your application owns its interface and product workflow. The kernel can also be hosted behind OUP, the Octos UI Protocol, so the same runtime can be controlled from a native app, a terminal, a browser, or another agent. Example applications built on it include the Octoscode terminal client and Octoscode Web.
The project also ships the applications that prove the kernel out. Octoscode is a terminal coding client built on Octos; on first launch it provisions a compatible local runtime automatically, so end users never interact with the kernel directly. A web variant offers the same agent through a browser. For builders, documentation describes how to embed the kernel, define tools and skills, and expose the runtime through the OUP protocol so custom interfaces — desktop, mobile, or automated controllers — can drive it.
Octos is free and open source under the Apache-2.0 license. There is no hosted tier or license fee; you build and run it as part of your own software.
Octos is for developers building their own AI agent products — coding tools, assistants, or automation apps — who want a proven runtime kernel instead of hand-rolling agent loops, memory, and tool plumbing.