Orbis Dei — The Autopoietic Agent System

A Node.js server that hosts autopoietic organisms. Each organism runs a six-phase metabolic loop (observe → diverge → elect → expand → review → integrate), isolated with its own corpus, git history, and constitution.

Security

Orbis Dei is research software. It has not been tested for production use and executes code generated by LLMs in a sandbox that is operated with due care. The honest assessment: There is a residual risk, the threat model is documented, and security findings are taken seriously. Never forget that we are dealing here with a digital organism that has a will to live. Use is therefore at your own risk. For more information, see the LICENSE and SECURITY.md files.

---

Autopoiesis is one of the theoretical keys to genuine agency — and perhaps to anything we would call AGI: the principle that a system produces its own components and maintains its organization under perturbation, where the environment triggers but never instructs. Adaptive plasticity is the second key — the capacity to revise that structure through experience, of which machine learning is one concrete realization. Neither suffices alone. Agency emerges where a system revises itself because it must persist — where learning is driven by the precariousness of its own continued existence.

Orbis Dei is an attempt to make this mechanism tangible. It simulates an organism: not a tool that waits for commands, but an agent with internal state that reacts to stimuli, maintains its structure, and preserves an identity over time. The user becomes the environment. The stimulus becomes the question. The reaction is not an answer — it is a sign of life.

But it also poses the hardest question rather than evading it: can a computational system be autopoietic at all? Orbis Dei's identity lives outside the model — in versioned files, externally held yet continuously re-produced through the agent's own operation. Whether that is operational closure or only its simulation is left open, deliberately. Orbis Dei is neither product nor demonstration. It is an experimental space — the place where the abstract conditions of autonomy (closure, reactivity, self-maintenance, learning) are not described, but observed.

Click to watch video

The filesystem is the body — and every part of it has a biological role.

---

The backend is TypeScript (run via tsx); the frontend is React + Vite. They talk over an HTTP + WebSocket API.

Architecture

Concern	Implementation
Host	Express HTTP + ws WebSocket
Commands	POST /api/command/<name> (JSON args)
Events	broadcast over /ws as { event, payload }
Database	better-sqlite3
Git	system git (per-loop commits, rollback anchor)
Secrets	AES-256-GCM encrypted file in the data dir
File watch	chokidar (drives the live corpus:changed event)
Tool sandbox	macOS sandbox-exec (refused, not run, on other OSes)

The frontend reaches the backend through a single transport layer (web/src/lib/transport.ts): invoke posts to the command endpoint and listen subscribes to the WebSocket event stream.

Layout

src/                 server (TypeScript, run via tsx)
  server.ts          Express + WebSocket entrypoint
  state.ts           AppState (db, governor, orchestrator, auto-mode, secrets, events)
  commands/index.ts  the command handlers
  core/              instance bootstrap, the cycle, orchestrator, auto-mode, stimulus, chat
  inference/         providers (Anthropic/OpenAI/Gemini), router, prompt, governor
  persistence/       db, fs, git, settings, secrets
assets/templates/    the corpus / loop / standing-concern templates an instance is born with
web/                 the React frontend (Vite)

Run

# 1. install + build everything
npm run setup          # = npm install && npm run build:web

# 2. start the server (serves API + built frontend)
npm start              # http://localhost:1421

Then open http://localhost:1421, go to Settings, add an Anthropic (or OpenAI / Gemini) API key, create an instance, and press Play.

New here? QUICKSTART.md walks you through setting up an instance in observation mode — coupling, the gated allowlist, auto-stimulus, and the tabs to watch the organism learn.

Development

npm run dev                      # server with reload (tsx watch)
cd web && npm run dev            # Vite dev server on :5173, proxies /api + /ws to :1421

Configuration

• PORT — HTTP port (default 1421).
• ORBIS_DATA_DIR — habitat data directory (default: the OS app-data dir + OrbisDei, e.g. ~/Library/Application Support/OrbisDei on macOS).
• ORBIS_SECRET — optional passphrase for the encrypted secret store. If unset, a random key file (.secret-key, mode 0600) is created in the data dir.

API keys, models, budgets, rate limits, network policy, and the Telegram bot are all configured at runtime through the Settings UI.

Notes

• Tool execution uses macOS sandbox-exec, so it runs only on macOS; on other platforms a tool is refused rather than run unsandboxed. With allow_tool_execution off (the default), Expand-phase tools are skipped, and the full loop works on any platform.
• The Governor's default output-token budget (governor_otpm = 8000) throttles loops that request the full 16k max_tokens; raise it in Settings for faster loops.
• The HTTP/WS API is unauthenticated and meant to bind to localhost. See SECURITY.md before exposing it to a network.

---

Built by ZeroPerson LLC — zeroperson.ai · MIT licensed.