cyber-ai-team

# Cyber AI Team [](https://github.com/Hyper-AI-Lab/cyber-ai-team/actions/workflows/ci.yml) [](LICENSE) [](docker-compose.yml) [](docs/architecture/autonomous-company-os.md)  **Cyber AI Team is a self-hosted AI company operating system for solo founders, one-person companies, and small digital startups.** It runs a governed team of AI workers around a human owner: a Chief Operating Agent, Observer Agent, company builder, memory steward, finance/accounting, legal, sales, marketing, support, product, engineering, operations, security, research, and communications roles. The goal is not another chatbot. The goal is a practical open-source foundation for an autonomous digital company: agents can observe company state, recall long-term memory, update canonical business records, propose or create roles, run safe internal actions, and escalate high-impact decisions to the owner. ## The Problem Solo founders and tiny teams increasingly need the operating surface of a real company: CRM, accounting, project management, customer support, knowledge work, compliance, security, outreach, monitoring, and decision logs. Hiring a full team is expensive, but gluing together many disconnected AI tools creates fragile context, duplicated work, and invisible risk. Cyber AI Team takes the opposite path: - **One company memory**, not scattered chat histories. - **One owner console**, not hidden autonomous actions. - **One canonical business record**, powered by ERPNext and PostgreSQL. - **Many specialized AI workers**, coordinated by a governed executive loop. - **Aggressive autonomy below impact thresholds**, with approval gates for risky or external actions. - **FOSS-first and self-hosted**, so early-stage founders are not forced into paid SaaS before revenue. This category is emerging quickly. [Business Insider reported](https://www.businessinsider.com/ai-agents-one-person-companies-china-openclaw-alibaba-president-2026-3) that AI agents are helping drive the rise of "one-person companies" in China, with Alibaba.com leadership estimating that 30-40% of its platform customers are solo entrepreneurs and describing agents as employees for those founders. [PwC's 2025 agentic AI survey](https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-agent-survey.html) found that 79% of surveyed executives were already adopting AI agents and 66% of adopters reported measurable productivity value. Cyber AI Team turns that idea into a self-hosted, inspectable operating system. ## What It Does | Capability | What Cyber AI Team provides | | --- | --- | | Autonomous company orchestration | Chief Operating Agent observes readiness, ERPNext, memory, agents, workflows, approvals, role gaps, KPIs, benchmarks, and owner instructions. | | Independent critique | Observer Agent reviews governor decisions for weak evidence, goal drift, prompt-injection-style instructions, policy violations, and unsafe assumptions. | | Dynamic role creation | Company Builder and role backlog detect missing business capabilities and create safe roles or request approval for risky ones. | | Long-term memory | Four-layer memory fabric: pinned identity, workflow state, retrieval memory with Qdrant, and canonical records in PostgreSQL/ERPNext. | | Business system of record | ERPNext integration for CRM, accounting, projects, tasks, tickets, procurement, customers, suppliers, leads, opportunities, and invoices. | | Human-in-the-loop governance | Owner approvals, replay protection, target matching, expiry, consumed-state checks, audit evidence, and readiness gates. | | Owner console | Next.js console for agents, memory, workflows, chat, approvals, integrations, audit trail, operations readiness, executive cockpit, role backlog, and governor state. | | Production-shaped operations | Docker Compose deployment, Caddy staging edge, migrations, backup/restore drills, smoke tests, load gate, CI, Trivy image scanning, Prometheus/Grafana-ready observability. | | FOSS-first resource policy | New tools and dependencies must declare license, cost, self-hostability, hosted-service dependency, data-sharing risk, and free-tier limitations. | ## Architecture ```mermaid flowchart TD Owner["Human owner"] --> Console["Owner Console"] Console --> API["FastAPI control plane"] API --> Governor["Chief Operating Agent"] API --> Observer["Observer Agent"] Governor --> Planner["Autonomous planning"] Governor --> Roles["AI role team"] Observer --> Governor Roles --> Tools["Governed tools"] Tools --> ERP["ERPNext canonical business records"] Tools --> Comms["Email and optional communications"] API --> Memory["Memory fabric"] Memory --> Qdrant["Qdrant retrieval"] Memory --> Postgres["PostgreSQL system state"] Planner --> Temporal["Temporal workflows"] API --> Audit["Audit and control evidence"] API --> Metrics["Metrics and readiness"] ``` Core stack: | Layer | Technology | | --- | --- | | Backend | FastAPI, SQLAlchemy, Alembic, PostgreSQL, Redis | | Agent orchestration | LangGraph/CrewAI-inspired role orchestration, governed planning services, Temporal workflows | | Memory | PostgreSQL, Qdrant, structured memory protocol, operation graph indexing | | System of record | ERPNext plus PostgreSQL | | Frontend | Next.js, React, Tailwind CSS | | Governance | Owner JWT auth, local/OPA authorization, approval records, audit/control evidence | | Observability | `/health`, `/ready`, `/metrics`, Prometheus/Grafana-compatible configs | | Deployment | Docker Compose, Caddy edge, staging promotion scripts, backup/restore runbooks | | Interoperability direction | MCP-style tool contracts and A2A-style agent interoperability surfaces | ## Who This Is For - Solo founders building AI-native businesses. - One-person companies that want digital workers without hiring a full staff. - Small startups that need operational leverage before they can afford a larger team. - Developers researching multi-agent systems, AI employees, autonomous agents, memory layers, ERP-backed agents, and human-in-the-loop AI governance. - Teams that want a self-hosted alternative to opaque agent SaaS. ## Quick Start ### 1. Configure ```bash cp .env.example .env # Edit .env. At minimum set owner credentials and your LLM provider key. ``` For production-like operation, replace all default secrets, set `OWNER_PASSWORD_HASH`, configure exact CORS origins, keep external side effects approval-gated, and validate integrations before enabling them. ### 2. Start ```bash # Recommended: managed screen session ./start.sh screen -r cyber-team # Or direct Docker Compose docker compose up --build ``` ### 3. Open | Service | Local URL | | --- | --- | | Owner Console | http://localhost:3001 | | API | http://localhost:8000 | | API docs | http://localhost:8000/docs | Sign in with `OWNER_EMAIL` and `OWNER_PASSWORD` from `.env`. ### 4. Build Your AI Company 1. Open the Owner Console. 2. Go to **Agents** and run the Company Builder. 3. Sync ERPNext company context if ERPNext is configured. 4. Review recommended roles in the role backlog. 5. Approve high-risk role/tool grants only when the payload and target are correct. ## Main AI Roles Cyber AI Team includes role families for: - Chief Operating Agent / governor - Observer Agent - Company Builder - Supervisor / orchestrator - Memory Steward - Finance and accounting - Legal and policy - Sales and CRM - Marketing and PR - Customer support and success - Product and project management - Software engineering and QA - Operations and procurement - People and HR - Security and compliance - Knowledge and research - Communications New roles can be proposed from company context, ERPNext drift, blocked work, role gaps, workflow failures, memory findings, or owner instructions. Low-risk internal roles can be created automatically; side-effectful or high-risk capabilities require owner approval. ## Safety Model Cyber AI Team is designed for autonomy with explicit control boundaries: - No fake-success tool execution. - No external side effects without matching approval in staging/production policy. - No approval replay, wrong-target approval, expired approval, or consumed approval execution. - No generated-code hot loading into the live runtime. - No paid/SaaS-only tool dependency as a readiness blocker under the FOSS-first policy. - No deletion of audit evidence through normal APIs. - Prompt-injection-style requests are downgraded to review instead of executed directly. The owner can inspect decisions, approvals, memory traces, role gaps, tool readiness, workflows, audit events, and governor/observer state from the console. ## Development ```bash # Backend cd backend pip install -e ".[dev]" pytest # Frontend cd frontend npm install npm run dev ``` Run the local production-readiness gate before opening a pull request: ```bash ./scripts/quality-gate.sh ``` Heavier checks are available: ```bash RUN_MIGRATION_REHEARSAL=1 RUN_COMPOSE_SMOKE=1 ./scripts/quality-gate.sh ``` ## Documentation - [Autonomous Company OS architecture](docs/architecture/autonomous-company-os.md) - [Production readiness plan](docs/production-readiness-plan.md) - [ERPNext runbook](docs/runbooks/erpnext.md) - [Backup and restore runbook](docs/runbooks/backup-restore.md) - [Deployment promotion runbook](docs/runbooks/deployment-promotion.md) - [Staging promotion runbook](docs/runbooks/staging-promotion.md) - [Release rollback runbook](docs/runbooks/release-rollback.md) - [GitHub discoverability notes](docs/visibility/github-discoverability.md) ## Repository Keywords AI agents, AI agent, agentic AI, autonomous agents, multi-agent system, AI employees, digital workers, one-person company, solo founder, startup automation, company OS, autonomous company, ERPNext, LangGraph, CrewAI, Temporal, Qdrant, FastAPI, Next.js, Docker Compose, MCP, A2A, RAG, long-term memory, human-in-the-loop, owner console. ## Status This is an active open-source project and a production-shaped reference implementation. The system already includes authenticated owner flows, ERPNext-backed operations, governed role backlog, readiness evidence, CI, staging deployment scripts, backup/restore drills, and safety gates. It is still evolving quickly, especially around the executive governor, observer critiques, adaptive benchmarks, and richer business integrations. ## Contributing Contributions are welcome if they preserve the core principles: self-hosted, FOSS-first, auditable, owner-visible, approval-gated for high impact, and honest about unavailable tools. Start with [CONTRIBUTING.md](CONTRIBUTING.md). ## License MIT. See [LICENSE](LICENSE).