octos-org/octos

Octos - Agentic Operating Systems

Octos – an Open‑Source “Agentic OS” you run yourself

What it is – Octos is a Rust‑written backend that lets you run a full‑featured AI‑assistant on your own machine (or on a self‑hosted server). It talks to any of the major LLM providers (OpenAI, Anthropic, Gemini, DeepSeek, etc.) and exposes a rich set of REST endpoints and a JSON‑RPC “UI Protocol” that client apps can use.

Why it matters – Most chat‑bot projects are single‑user, single‑model services. Octos is built for multi‑tenant use: a single 31 MB binary can host hundreds of isolated “profiles”, each with its own prompt, model, tools, and memory. It also supports multi‑LLM pipelines (different models for different steps) and swarm dispatch, so you can fan‑out work to many workers and aggregate the results.

Key components

  • Kernel (this repo) – the core runtime: provider adapters, tool sandbox, memory layers, session handling, and the HTTP/WebSocket API.
  • octos‑web – a browser‑based UI (chat, voice, document creation, admin dashboard) that is bundled into the binary and served at /app/.
  • octoscode – a terminal‑style client similar to Claude Code.

How you get it running

  1. Install the binary via Homebrew, npm, or the provided install.sh script.
  2. Run octos init to pick a provider and model, then log in with octos auth login.
  3. Start the server with octos serve --solo (local dev) or install it as a background service (port 8080).
  4. Open http://localhost:50080/app/ (or the service port) and start chatting.

Features you can actually use

  • 80+ REST endpoints for chat, sessions, profiles, skills, scheduling, metrics, webhooks, etc.
  • 14 messaging channels (Telegram, Discord, WhatsApp, Slack‑like services, Matrix, etc.) – the same backend can be reached from many chat apps.
  • Tool sandbox – safe execution of code, web‑search, file manipulation, office‑suite generation (DOCX/PPTX/XLSX) using bwrap, Landlock, Docker, or Windows AppContainer.
  • Three‑layer memory – long‑term entity store, episodic Redb store, and per‑session JSONL with LLM‑driven compaction.
  • Autonomy primitives – /goal <objective> keeps the agent working across turns, /loop runs periodic tasks, and you can roll back a session to any previous turn.
  • Swarm dispatcher – fan‑out a request to many workers (including external MCP tools) and combine results, with built‑in cost tracking and fail‑over.
  • Provider fail‑over – a three‑layer router (Retry → Chain → Adaptive) that can hedge between models and circuit‑break on errors.
  • Bindings for other languages – C/Go (octos-ffi), Python (octos-pyo3), Swift/Kotlin (octos-uniffi), and WebAssembly (octos-wasm).
  • UI Protocol v1 – JSON‑RPC over WebSocket or stdio, enabling any front‑end (web, terminal, editor) to talk to the same backend.

Typical use‑cases

  • A personal AI assistant that can run code, browse the web, and keep notes, all while keeping your data local.
  • A team‑wide “agentic OS” where each user gets their own profile but shares the same server.
  • Embedding the agent loop inside other software via the Rust crates or the language bindings.
  • Building custom agents that call external tools through the MCP (Model Client Protocol) system.

Where to learn more

Bottom line – Octos is a genuine, production‑grade open‑source platform for running AI agents locally or self‑hosted, with a focus on multi‑tenant isolation, extensible tooling, and deep integration with many LLM providers. It is well‑documented, actively maintained, and designed to be the “operating system” for AI‑driven applications.

Related

  • Project
    moltis-org/moltis

    A secure, persistent personal agent server written in Rust that provides sandboxed tool execution and multi-channel AI interaction.

  • Project
    jaylfc/taOS

    taOS is an open‑source, self‑hosted AI platform that provides a browser‑based desktop, a distributed compute cluster, and a local memory service (taOSmd) for AI agents. It runs on any hardware (PCs, SBCs, phones), lets you install and manage agents from 17 frameworks, and keeps all conversations, files and knowledge graphs on your own machines, with cloud models optional.

  • Project
    holaboss-ai/holaOS

    An agentic workspace that integrates AI agents side-by-side with real applications and chat tools, featuring local-first memory and shared tool access.

  • Project
    AgentsMesh/AgentsMesh

    An AI agent workforce platform that allows a single operator to deploy, monitor, and coordinate hundreds of isolated AI agents across a fleet of self-hosted runners.

  • Project
    Mng-dev-ai/agentrove

    A self-hosted AI coding workspace that orchestrates multiple AI agents from one interface, providing isolated sandboxes and multi-agent workflows.