OpenCognit Documentation
The orchestration platform for autonomous AI agents — the operating system for your Zero-Human Company.
Open Source · Self-Hosted · Zero Human Company
Welcome to OpenCognit
OpenCognit is a multi-agent orchestration platform that enables specialized AI agents to work together autonomously — like a real company operating around the clock.
Quick Start
bash <(curl -fsSL https://raw.githubusercontent.com/OpenCognit/opencognit/main/install.sh)One command. Clones the repo, installs deps, generates secrets, migrates the DB. Then cd my-opencognit && npm run dev.
Getting Started
Install, configure and launch your first company in 5 minutes.
Configuration
Configure API keys, LLM providers and agent settings.
Architecture
Understand the CEO-Expert model and the Heartbeat engine.
Experts (Agents)
Create agents, choose adapters and assign skills.
What is OpenCognit?
OpenCognit implements the Zero-Human Company concept: instead of using AI as a simple chat tool, specialized agents are organized into roles (CEO, Developer, Researcher) that independently take on tasks, produce results and coordinate with each other.
The Core Model
| Role | Description |
|---|---|
| CEO Agent | Breaks down strategic goals into tasks and delegates to experts |
| Experts | Specialized agents with skills (Dev, Research, DevOps, etc.) |
| Heartbeat | Autonomous timer — agents work without human prompts |
| Approvals | Humans retain control over critical decisions |
Supported LLM Providers
Claude (Anthropic)
Claude API directly or via Claude Code CLI.
OpenRouter
Access to 100+ models (GPT-4, Gemini, Mistral, etc.).
Ollama
Local open-source models without cloud dependency.
Bash & HTTP
Shell commands and HTTP APIs as adapters.
Feature Overview
- CEO Orchestrator — Extended Thinking: sets goals, delegates, reviews, escalates — fully autonomous
- Persistent Memory — Every agent remembers everything across sessions (Rooms, Diary, Knowledge Graph)
- Real Execution — HTTP, Bash, Claude Code CLI, external APIs — agents act, not just reason
- Self-Healing Retries — Automatic retry with backoff (5→15→45 min) on failure
- Orchestrator Escalation — After 3 failures, an escalation task is auto-created for the CEO
- Critic/Evaluator Loop — Every output is reviewed before a task is marked done; after 2 rejections the task is set to
blockedfor human review - Dual-Engine Quality Gate — Both heartbeat and scheduler execution paths pass through the Critic
- Task Chaining — When Task A completes, dependent tasks unlock and agents wake automatically
- Atomic Budget Control — Cent-precise limits per agent, hard stop on budget exceeded
- Approval Workflows — Hiring, strategy and budget require board approval
- Routines & Cron — Scheduled tasks with 5-field cron expressions
- SOUL.md Identity — File-based agent personality, git-trackable
- LRU Conversation Memory — Conversation history uses true last-access-time eviction
- CLI Lock Safety — Separate agent/chat locks with 5-min timeout prevent Telegram from blocking agents
- DB Performance Indexes — 9 indexes on hot-query columns for fast inbox and status lookups
- Real-Time Dashboard — WebSocket live updates, Kanban, War Room, activity log
- Self-Hosted — SQLite backend, no external database, no cloud dependency