Open Cognitive Tasks Orchestration System
One brain, many arms
Like an octopus — eight arms thinking in parallel, running async, self-healing, self-evolving. Multi-agent collaboration, multi-step workflows, multi-model failover — runs everywhere: phones, PCs, servers, Raspberry Pi, RISC-V. Zero dependencies.
Capabilities
Why Octos
Your agents, your way
92 API endpoints to build any interface on top — web dashboards, mobile apps, CLI tools, or embed AI into your existing product.
Room for everyone
One 31MB binary serves 200+ users on a single machine. Each person gets their own agent with private memory, sessions, and data.
Best model for each task
Cheap models for planning, powerful ones for synthesis. Chain multiple LLMs in a single workflow with automatic failover when one goes down.
Always online
Three layers of failover keep your agents running even when a provider goes down. Hedge racing, circuit breakers, automatic switching.
Smart tool management
30+ tools available, but only 15 loaded at once. Octos activates what your agent needs and quietly puts away the rest — fast responses, no bloat.
Conversations that flow
Five queue modes control how your agent handles busy moments — queue messages, collect them, steer mid-conversation, or let users interrupt.
It remembers
Your agent knows what you talked about yesterday, your preferences, your projects. Three memory layers keep the right context in reach.
Documents, slides, spreadsheets
Generate PPTX, DOCX, and XLSX natively — no LibreOffice, no external tools, no network calls. Pure Rust, zero dependencies.
Built safe
Agents run in isolated sandboxes. No unsafe code anywhere in the project. 67 prompt injection tests and counting.
Architecture
Simple inside, powerful outside
Eight focused modules, one 31MB binary. No runtime, no Docker, no pip install. Download and run.
connects 12 messaging platforms
thinks, uses tools, stays safe
runs multi-step workflows
talks to 15 AI providers
never forgets what matters
Deep Research
Think deeper, search wider
One agent designs the research plan, others fan out to search in parallel. Cheap models plan, powerful ones synthesize. Every step is a different model.
8 workers at once
Search workers fan out in parallel, then converge on a single answer.
5 search providers
Tavily, DuckDuckGo, Brave, You.com, Perplexity — if one is slow, the others finish first.
Self-healing routes
When a provider slows down, Octos notices and routes around it automatically.
Tools
A well-organized toolbox
30+ tools available, but your agent only loads what it needs right now. Idle tools quietly step aside — keeping responses fast and context windows lean.
Active tools
Always ready, instant response
On standby
Activated when needed, put away when done
Catalog
No limit on what your agent can learn to do
11 core tools
12 agent tools
7 app skills
Memory
It remembers
Your agent knows what you talked about yesterday. It remembers your preferences, your projects, your decisions — and forgets nothing you want kept. Three layers, one seamless experience.
Long-term
What it always knows about you
Preferences, projects, recurring topics — auto-extracted and ready for instant recall.
Episodic
What it learned from each task
Files modified, decisions made, outcomes achieved — searchable by keyword or meaning.
Session
The current conversation
Always in context, auto-compacted when it gets long. Switch threads with /new, /s, /back.
Content
Documents, slides, spreadsheets
Ask your agent to make a slide deck, write a report, or build a spreadsheet. It generates real Office files — no LibreOffice, no network calls, zero dependencies.
Slide decks with layouts, images, and charts
Documents with headings, tables, and images
Spreadsheets with formulas, styling, and charts
AI Art Pipeline
Many visual styles
Cards, comics, infographics, posters, diagrams, timelines, and more. Each style is a composable template with AI-driven content placement.
Security
Built to be trusted
Your data stays on your machine. Every agent runs in a sandbox. No unsafe code in the entire project. Security isn't an afterthought — it's the foundation.
Agents can't break out
Each agent runs in its own isolated sandbox that blocks dangerous operations — on Linux, macOS, and Docker.
No unsafe code
The entire codebase enforces deny(unsafe_code). Pure Rust with rustls for TLS — no OpenSSL, no C dependencies.
Prompt injection defense
67 injection tests covering input sanitization and output filtering. Constant-time comparison prevents timing attacks.
Your data stays yours
Each agent operates in a scoped workspace. File access is restricted to its own directory. No cross-tenant data leakage.
Safe credentials
macOS Keychain integration for API key storage. No plaintext secrets in config files.
Process isolation
Each user gets a separate OS process with its own memory space, sessions, and data — true multi-tenancy.
Performance
Run on Your Spared Computer
200+ AI agents serving thousands of users — on hardware you already own. No Kubernetes, no auto-scaling, no 3am pages. 20MB per agent, cold start under 100ms.
| Metric | Octos | Others | Note |
|---|---|---|---|
| Startup time | <100ms | 500ms–2s | Cold start to ready |
| Memory per agent | 20MB | 200–300MB | Baseline resident |
| Memory per session | Few KB | 10–50MB | Green threads |
| Binary size | 31MB | 200MB+ | Static vs node_modules |
| Agents per machine | 200+ | ~20 | On a 16GB Mac Mini |
| GC pauses | None | Unpredictable | Deterministic latency |
| Parallelism | True | Event loop | Multi-core utilization |
| Search workers | 8 concurrent | Sequential | Deep research tasks |
Gateway
Meet your users where they are
Your AI lives on Telegram, Discord, Slack, WhatsApp, WeChat, and 7 more platforms. One agent, every channel — with full conversation history wherever your users talk.
LLM Layer
Use Your Favorite Tokens
Pick the right model for each task. Chain providers in a single pipeline — if one goes down, the next one takes over automatically.
Comparison
Different animals, different oceans
Not better or worse — just different. Octos is infrastructure, not a library. If you need a chatbot SDK, LangChain is great. If you need an AI operating system, keep reading.
| Dimension | Octos | OpenClaw |
|---|---|---|
| What it is | AI operating system | Chatbot app |
| Install size | 31MB | 200MB+ |
| Startup | <100ms | 500ms-2s |
| Memory per agent | 20MB | 200-300MB |
| Agents per machine | 200+ | ~20 |
| LLM providers | 15 built-in | 10+ |
| Channels | 12 | 22+ |
| Failover | 3-layer auto | Basic retry |
| Tool loading | Smart (LRU) | All upfront |
| Document generation | PPTX/DOCX/XLSX | None |
| Voice | ASR + TTS | Wake word + TTS |
| Memory | 3-layer hybrid | SQLite + vector |
| Security | deny(unsafe_code) | Docker sandbox |
Use Cases
Built for real workloads
AI in your product
Give every customer their own AI assistant — isolated, personalized, always available. One binary handles them all through a clean REST API.
Customer support everywhere
One agent on Telegram, Discord, Slack, WhatsApp, Email — everywhere your users are. They switch channels without losing context.
Deep research pipelines
Let cheap models plan the search, powerful ones synthesize the answer. 8 workers search in parallel across 5 providers.
Content at scale
Generate slides, documents, and spreadsheets natively. AI art pipeline with 17 styles for cards, comics, infographics, and posters.
AI for the whole team
200+ specialized agents on a single machine. Sub-accounts for family plans, isolated workspaces for each team member.
Runs anywhere
31MB binary, cold start under 100ms, zero garbage collection pauses. Deploy on a Mac Mini, a cloud VM, or a Raspberry Pi.
Get Started
Deploy in minutes
No Python, no Docker, no runtime dependencies. Install the binary, set your API key, and start talking to your AI. That's the whole process.
# Install
cargo install --path crates/octos-cli
# Initialize configuration
octos init
# Set your API key
export ANTHROPIC_API_KEY=your-key
# Start chatting in the terminal
octos chat
# Or launch the full gateway
octos gateway
# Or start the REST API server
octos serveQuestions
FAQ
How is Octos different from LangChain or CrewAI? +
Octos is an operating system, not a framework. It runs as a single binary — no Python, no Docker, no glue code. Built-in multi-tenancy, 12 messaging channels, and automatic provider failover. You configure agents through a REST API or a JSON file.
What LLMs does Octos support? +
15 providers out of the box: Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter, and more. You can chain multiple providers in a single pipeline — cheap models for planning, powerful ones for synthesis — with automatic failover if one goes down.
Can I run Octos on a small server? +
Yes. The binary is 31MB, cold starts in under 100ms, and needs zero external dependencies. A single Mac Mini can run 200+ agents. No Python runtime, no Docker, no Kubernetes.
Is Octos open source? +
Yes. Octos is fully open source under the MIT license and available on GitHub.
How do I connect a messaging channel? +
Just add the channel config to your profile. Telegram, Discord, Slack, WhatsApp, WeChat, Email, and 6 more are built in — no plugins needed. Your agent shows up on all of them simultaneously.
Can multiple users share one Octos instance? +
Absolutely. Each user gets their own isolated agent with separate memory, sessions, and data. A family plan mode even supports sub-accounts — parent manages the billing, each person gets their own AI.
Deploy your first agent
in 5 minutes
Download. Set your API key. Talk to your AI on any channel. That's it.