Etz Chaim AI

A cognitive operating system for LLM agents. Apache 2.0 · Multi-provider · Standards-first · Anthropic-optimized

In the SOAR / ACT-R / CLARION / LIDA lineage, Etz Chaim AI gives LLM agents 10 cognitive faculties, 13 rectifiers, 11 adversarial probes, 22 typed paths, 1696 primary-source specs with E1–E6 confidence labels, and a nightly auto-improve daemon that mutates its own specifications based on observed failures.

It works on Anthropic Claude, OpenAI GPT-5.5, Google Gemini 3, and local models via Ollama — switch in <30 minutes. See PORTABILITY.md.

Today's coding agents are powerful but ahistorical. They forget. They repeat the same mistakes. They have no model of what they don't know, no way to detect when they're drifting from spec, no nightly housekeeping that consolidates yesterday's failures into tomorrow's improvements.

Etz Chaim AI fills that gap with composable cognitive primitives that plug into Claude Code, Codex CLI, Cursor, Windsurf, Antigravity, OpenCode, or your own custom harness via the open Agent Skills + MCP standards.

We support three install tiers based on your budget and trust model.

Fully local. No external API calls. Privacy-first.

# Prerequisites: Python 3.12+, Docker (for pgvector + TimescaleDB)
# Recommended local model: qwen3:72b or llama3.3:70b
ollama pull qwen3:72b

pipx install etzchaim
etzchaim init --provider ollama --model qwen3:72b
etzchaim doctor    # validates the install

Use cases: privacy-sensitive research, education, OSS contributors, air-gapped. Trade-offs: slower inference; smaller context; some features (Auto Mode, Channels, Routines, Managed Agents) unavailable.

API key from any provider. Most flexible.

pipx install etzchaim
export OPENAI_API_KEY=sk-...          # or ANTHROPIC_API_KEY or GEMINI_API_KEY
etzchaim init --provider openai --model gpt-5.5
etzchaim doctor

Supported providers (via LiteLLM, 100+): Anthropic, OpenAI, Google Vertex / Gemini, AWS Bedrock, Azure OpenAI, xAI, Mistral, Cohere, DeepSeek, Groq, Together, Fireworks, OpenRouter, Perplexity, NVIDIA NIM, Cloudflare AI, Replicate, vLLM, LM Studio.

Use cases: most contributors; teams evaluating Etz Chaim; multi-provider experiments.

Full Etz Chaim experience with the Anthropic-bonus layer wired in.

# Open in VS Code or GitHub Codespaces and "Reopen in Container"
gh repo clone yohanpoul/etz-chaim-ai
code etz-chaim-ai
# F1 → "Dev Containers: Reopen in Container"
# (or open in Codespaces — free 60h/month tier works)

Unlocks:

• Auto Mode (Max/Team/Enterprise) — Sonnet 4.6 classifier on every tool call
• Channels — Telegram / Discord / iMessage push alerts for drift events
• Claude Code Routines — nightly auto-improve loop runs on Anthropic infrastructure (your laptop doesn't need to be on)
• Managed Agents + Dreaming + Outcomes + Multiagent orchestration — the 11 malakhim adversaries as parallel specialist sub-agents with persistent memory
• Cowork Dispatch — assign work to Etz Chaim from your phone, work finishes on your desktop

If you maintain an active OSS project, apply for the Claude for Open Source Program — 6 months of Claude Max free for approved OSS maintainers. Deadline: June 30, 2026. See scripts/apply-oss-program.md for our application template.

After install, in your project directory:

# 1. Run the doctor
etzchaim doctor               # 20 health checks should pass

# 2. Try a faculty
etzchaim faculty causal-pearl \
  --query "Does drinking coffee cause better focus?"

# 3. Run the bidirectional spec↔code audit
etzchaim verify-bidirectional

# 4. (Optional) trigger the auto-improve loop manually
etzchaim improve --once

# 5. Connect a coding agent
# Claude Code: just run `claude` in this directory; .claude/ and skills/ are pre-wired
# Codex CLI:   run `codex` — reads AGENTS.md and .codex/skills/
# Cursor:      open the folder — .cursor/rules/etz-base.mdc references AGENTS.md

┌──────────────────────────────────────────────────────────────────┐
│  10 facultés cognitives                                          │
│  exploration · judgment · causal · dissensus · insight           │
│  selfmodel · selfmap · epistememory · failuretoinsight · intent  │
└─────────────────────┬────────────────────────────────────────────┘
                      │ compose via 22 sentiers (typed paths)
                      ▼
┌──────────────────────────────────────────────────────────────────┐
│  6 mature configurations                                         │
│  diagnose · explore · judge · synthesize · audit · learn         │
└─────────────────────┬────────────────────────────────────────────┘
                      │ policed by
                      ▼
┌──────────────────────────────────────────────────────────────────┐
│  13 rectifiers + 11 malakhim adversaries                         │
│  Background daemons detecting drift, contradictions, decay       │
└─────────────────────┬────────────────────────────────────────────┘
                      │ feed into
                      ▼
┌──────────────────────────────────────────────────────────────────┐
│  Nightly auto-improve daemon (Karpathy pattern)                  │
│  Read failure traces → propose spec mutations → verify → PR      │
└──────────────────────────────────────────────────────────────────┘

The full architecture: memory/ARCHITECTURE.md.

Etz Chaim is built on open standards, not vendor-specific APIs:

• Agent Skills — 13 portable SKILL.md files, work in Claude Code, Codex CLI, Cursor, Gemini CLI, Antigravity, OpenCode, goose, Letta, Amp, Devin (35+ platforms)
• Model Context Protocol — etzchaim-mcp server consumable by any MCP client (Claude Code, Codex, Cursor, Windsurf, ChatGPT Dev Mode)
• AGENTS.md (Linux Foundation AAIF) — single source of truth, symlinked to CLAUDE.md / .codex/AGENTS.md
• LiteLLM — 100+ LLM providers via OpenAI-compatible interface

Switching from Anthropic Claude to OpenAI GPT-5.5 is one line in litellm.config.yaml. The cognitive engine, the rectifiers, the spec corpus, the daemon — all keep working. See PORTABILITY.md.

Etz Chaim follows the Boris Cherny workflow adapted for cognitive-architecture work:

1. Start in Plan mode (Shift+Tab×2 in Claude Code) for any spec mutation
2. Run the verify-spec subagent before marking complete (the 2-3× quality boost insight)
3. Every mistake → memory/MISTAKES.md rule (/mistake-to-rule)
4. For batch changes → /batch-migration (parallel worktree workers)
5. For adversarial validation → /adversarial-probe (11 malakhim in parallel)
6. CLAUDE.md is a symlink to AGENTS.md — edit AGENTS.md only

Behavioral rules and Boris references in AGENTS.md. Full anti-pattern catalog in memory/MISTAKES.md.

See memory/DECISIONS.md for our Architecture Decision Records (ADRs) — particularly:

• ADR-0001: Standards-first architecture
• ADR-0003: Boris's verify-spec subagent pattern
• ADR-0004: Skills + 5 subagents (Barry Zhang's thesis applied)

• AGENTS.md — agent instructions (source unique)
• memory/ARCHITECTURE.md — system invariants & faculty contracts
• memory/MISTAKES.md — anti-patterns (living doc)
• memory/DECISIONS.md — ADRs
• docs/PORTABILITY.md — cross-provider switching
• docs/CODE_WITH_CLAUDE_2026.md — Anthropic 2026 features map
• docs/REFERENCES.md — full external reference index

Apache 2.0 — see LICENSE.

If you use Etz Chaim AI in research, please cite:

@software{etz_chaim_ai_2026,
  author = {Yohan Poul},
  title = {Etz Chaim AI: A Cognitive Operating System for LLM Agents},
  year = {2026},
  url = {https://github.com/yohanpoul/etz-chaim-ai},
  license = {Apache-2.0}
}

Etz Chaim AI builds directly on the work of:

• Anthropic for the Skills + MCP + Auto Mode foundations, and the engineering papers that shaped the architecture
• Boris Cherny (creator of Claude Code) whose workflow patterns are encoded throughout this project — thread
• Barry Zhang & Mahesh Murag for the "Don't build agents, build skills" thesis we apply
• Erik Schluntz for the Building Effective Agents patterns
• Andrej Karpathy for the autonomous research loop concept that underlies our nightly daemon
• Sun, Anderson, Newell, Franklin, Pearl for the cognitive architecture lineage Etz Chaim sits in
• The Linux Foundation AAIF for stewarding the AGENTS.md and MCP standards
• The agentskills.io community for the cross-tool skills standard
• The LiteLLM team for making provider-agnostic LLM dispatch routine

Full reference index: docs/REFERENCES.md.