I built an open-source AI agent OS

Build Your Zero Human Company. The open-source AI agent OS — CEO orchestrator, persistent memory, real execution, atomic budgets. Self-hosted. No cloud lock-in. Free forever. 🚀 Quick Start · 📖 Docs · 💬 Community · ☕ Support If OpenCognit saves you time, a ⭐ star helps others find it. OpenCognit is an AI agent orchestration OS — not a chatbot, not a single-agent wrapper. It runs a virtual company of autonomous AI agents that work together without you watching. You set a goal. The CEO agent breaks it down, assigns tasks to specialists, reviews their work with a built-in Critic loop, and escalates blockers — while you sleep. You → Goal → CEO Agent → Dev Agent → Writer Agent → Researcher Agent ↑ ↓ ↓ Persistent Memory ←── Critic ←── Results ←────┘ Yes, if: - You want AI agents that actually remember — not just across a conversation, but across days and weeks - Your CEO agent should reason about the company state before delegating, not just execute prompts - You need a Critic loop — no silent "done" without quality review - You want hard budget limits enforced at the cent level, not just estimates after the fact - You need an org chart, not a pipeline — agents with roles, hierarchies, and peer meetings - You're building something real and want the full control plane: Goals, Kanban, War Room, Activity Feed Probably not, if: - You just want to chain a few prompts together — use LangChain or a simple script - You need a no-code drag-and-drop workflow builder — try n8n - You only ever run one agent at a time | Problem | OpenCognit | |---|---| | Agents lose context on restart | Persistent Memory memory per agent | | No quality control — agents ship anything | Critic/Evaluator Loop: every output is reviewed before marking done | | API costs run away overnight | Atomic Budgets: agents pause at your limit | | Task context lost between agents | Task-Output-as-Input: blocker results flow downstream automatically | | Sequential agent wakeups (slow) | Parallel wakeups via Promise.all | | Manual task orchestration | CEO Extended Thinking: Sonnet reasons before acting | | No memory across sessions | SOUL documents + Memory per agent | | Scattered configs | Full UI: Org Chart, Goals, Kanban, War Room | The CEO agent uses claude-sonnet-4-6 with Anthropic's Extended Thinking — it reasons step-by-step before making decisions. It delegates tasks via native tool calls (not JSON parsing), calls team meetings, and requests board approval for hiring. Each agent has its own memory with Rooms (key-value), Diary (structured entries), and a Knowledge Graph (subject-predicate-object triplets). Agents tag important outputs with [REMEMBER:room] — parsed and saved automatically. Old memories consolidate into LLM-compressed summaries. Every task result goes through a lightweight evaluator (Haiku) before being marked done . If insufficient, the agent retries — up to 2 times — before escalating to the board. No silent failures. Define blocking relationships between tasks. When Task A completes, dependent tasks are automatically unlocked and their assigned agents are woken — no manual intervention. Each agent gets a structured identity document — IDENTITY, DECISION PRINCIPLES, CYCLE CHECKLIST, PERSONALITY — generated by LLM or hand-crafted. Solves the "Memento Problem" where agents forget who they are across sessions. Each agent wakes on its own schedule (5m, 15m, 1h, 24h, or custom). The heartbeat is the core of autonomy — it processes inbox tasks, runs adapters, tracks costs, and feeds results back to the team. - Dashboard — live agent grid, system pulse, company health score - War Room — full-screen mission control with animated counters - Meetings — agents can call peer meetings, synthesize results - Goals (OKR) — 4-level hierarchy, progress tracking - Kanban — tasks with Critic-reviewed outputs - Activity — 28-day heatmap, filtered feed - Routines — cron-based automations - Memory — browsable agent memory with Knowledge Graph viz - Telegram — mobile interface with /status ,/tasks , free-form chat - OpenClaw Gateway — bridge existing OpenClaw agents into your company with their full knowledge base bash <(curl -fsSL https://raw.githubusercontent.com/OpenCognit/opencognit/main/install.sh) Interactive wizard: asks for project name + optional API key, clones the repo, installs deps, generates secrets, migrates DB, builds frontend. Done in under 2 minutes. opencognit # → http://localhost:3201 git clone https://github.com/OpenCognit/opencognit.git cd opencognit docker compose up # → http://localhost:3201 Data is persisted in ./data/ — safe across restarts. Requirements: Node.js ≥ 20 git clone https://github.com/OpenCognit/opencognit.git cd opencognit bash setup.sh # installs deps, generates keys, migrates DB, builds frontend npm start # → http://localhost:3201 - Create your company — name it, describe its goal - Configure API keys — Settings → add Anthropic or OpenRouter key - Set default model — Settings → Standard-Modell (all agents use this unless overridden) - Add a CEO agent — set connection type to ceo , enable Orchestrator flag - Add worker agents — give them roles, skills, heartbeat intervals - Create your first goal — CEO will start breaking it down autonomously | Provider | Use Case | |---|---| | Anthropic (Claude) | CEO Extended Thinking, Critic Loop, SOUL generation | | OpenRouter | Worker agents (100+ models, unified API) | | OpenAI | GPT-4o, GPT-4o-mini | | Ollama | Local models (Llama 3, Mistral, etc.) | | Claude Code CLI | Code execution with session persistence | | Gemini CLI | Google models | | Bash | Shell commands | | HTTP | External APIs | | OpenClaw Gateway | Connect existing OpenClaw agents with their full knowledge base | ┌──────────────────────────────────────────────────────────────┐ │ React 19 Frontend │ │ Dashboard · Agents · Tasks · Meetings · Intelligence │ │ Goals · War Room · Memory (Palace) · Costs · SOUL Editor │ └───────────────────────────┬──────────────────────────────────┘ │ REST + WebSocket + SSE ┌───────────────────────────▼──────────────────────────────────┐ │ Express 5 API Server │ │ │ │ ┌────────────────────────────────────────────────────────┐ │ │ │ Heartbeat Engine (modular) │ │ │ │ context-builder · critic · budget · notifications │ │ │ │ actions-orchestrator · actions-worker · dependencies │ │ │ └────────────────────────────┬───────────────────────────┘ │ │ │ │ │ ┌──────────────┐ ┌──────────▼───────┐ ┌───────────────┐ │ │ │ Cron Sched │ │ Adapter Registry │ │ CEO Adapter │ │ │ │ (30s tick) │ │ Bash · HTTP │ │ + Thinking │ │ │ └──────────────┘ │ Claude Code │ └───────────────┘ │ │ │ OpenRouter · LLM │ │ │ │ OpenClaw Gateway │ │ │ └──────────┬────────┘ │ │ │ │ │ ┌────────────────────────────▼───────────────────────────┐ │ │ │ MemPalace (per agent) │ │ │ │ Wing · Rooms · Diary · Knowledge Graph · Summaries │ │ │ └────────────────────────────────────────────────────────┘ │ └──────────────────────────────────────────────────────────────┘ │ ┌───────────────────────────▼──────────────────────────────────┐ │ SQLite / PostgreSQL (self-hosted, no cloud) │ │ 35 tables: Agents · Tasks · Goals · Memory · Budget · etc. │ └──────────────────────────────────────────────────────────────┘ | Layer | Technology | |---|---| | Frontend | React 19 + Vite 6 | | UI | Glassmorphism design system, Lucide icons | | Routing | React Router v7 | | Backend | Express 5 (Node.js) | | Database | SQLite via better-sqlite3 + Drizzle ORM | | Auth | JWT + bcrypt | | Real-time | WebSocket + SSE | | Language | TypeScript (strict, full-stack) | | Runtimes | Claude Code CLI, Anthropic, OpenRouter, Ollama, OpenAI | opencognit/ ├── server/ │ ├── index.ts # Express + all API endpoints │ ├── db/ │ │ ├── schema.ts # Drizzle schema (SQLite, 35 tables) │ │ └── schema.pg.ts # Drizzle schema (PostgreSQL, full parity) │ ├── services/ │ │ ├── heartbeat/ # Modular heartbeat engine (10 files) │ │ │ ├── service.ts # HeartbeatServiceImpl (orchestrator) │ │ │ ├── critic.ts # Critic loop + advisor plan/correction │ │ │ ├── context-builder.ts # Memory/goal/workspace injection │ │ │ ├── actions-orchestrator.ts # CEO action dispatcher │ │ │ ├── actions-worker.ts # Worker bash/JSON actions │ │ │ ├── dependencies.ts # Task chaining + blocker scan │ │ │ ├── notifications.ts # CEO feedback loop + meetings │ │ │ ├── budget.ts # Budget enforcement │ │ │ └── utils.ts # Soul cache, trace, focus mode │ │ ├── cron.ts # Cron scheduler │ │ ├── wakeup.ts # Wakeup coalescing │ │ ├── memory-auto.ts # Auto-save + REMEMBER protocol │ │ └── messaging.ts # Telegram + channels │ ├── routes/ │ │ └── webhooks.ts # Telegram · WhatsApp · Slack · Routine webhooks │ └── adapters/ │ ├── ceo.ts # CEO Orchestrator + Extended Thinking │ ├── claude-code.ts # Claude Code CLI adapter │ └── registry.ts # Adapter auto-selection └── src/ ├── pages/ # All UI pages ├── components/ # Shared components └── i18n/ # DE + EN translations OpenCognit is 100% self-hosted. Your data never leaves your machine: - Database: SQLite file in ./data/ - API keys: encrypted at rest - No telemetry, no analytics, no external dependencies The only outbound connections are the LLM API calls you configure. - Electron desktop app (one-click, no terminal) - Agent marketplace (community plugins) - Multi-tenant cloud deployment option - Web-based IDE integration (Cursor / Windsurf) - Mobile dashboard - Agent OAuth delegation | OpenCognit | AutoGPT | CrewAI | LangChain Agents | n8n AI | | |---|---|---|---|---|---| | Multi-agent org chart | ✅ | ❌ | ✅ | ❌ | ❌ | | CEO orchestrator | ✅ | ❌ | ❌ | ❌ | ❌ | | Persistent per-agent memory | ✅ | ❌ | ❌ | ❌ | | | Built-in Critic/QA loop | ✅ | ❌ | ❌ | ❌ | ❌ | | Atomic budget per agent | ✅ | ❌ | ❌ | ❌ | ❌ | | Full UI (no code) | ✅ | ❌ | ❌ | ✅ | | | Self-hosted, no cloud | ✅ | ✅ | ✅ | ✅ | ✅ | | Local models (Ollama) | ✅ | ✅ | ✅ | ✅ | | | Claude Code CLI integration | ✅ | ❌ | ❌ | ❌ | ❌ | | OpenClaw agent bridge | ✅ | ❌ | ❌ | ❌ | ❌ | | Free forever | ✅ | ✅ | ✅ | ✅ | Looking for an AutoGPT alternative, CrewAI alternative, or open-source multi-agent system? OpenCognit adds a full org chart, persistent memory, and a built-in quality loop on top. Pull requests welcome. See CONTRIBUTING.md for guidelines. OpenCognit is free and open-source. If it saves you time or helps your business, consider supporting development: Every contribution helps keep this project alive and actively maintained. AGPL-3.0 — see LICENSE Commercial licensing (SaaS / enterprise closed-source) available on request via GitHub Issues.