JARVIS – Sistema Multi-Camadas para Claude Code
WGlynn/JARVIS
Arquitetura de agente resiliente com persistência e gates anti-alucinação para Claude
Ver no GitHubUm sistema de agente com oito camadas de arquitetura que adiciona persistência, anti-alucinação, hooks determinísticos e meta-protocolos sobre Claude Code. Inclui 47 hooks prontos, 468 primitivas markdown, corpus tipado e CLI para verificação e navegação de padrões.
Instalação
npx convez-ai WGlynn/JARVIS1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Construir agentes Claude Code com memória entre sessões
- Implementar gates determinísticos em tool calls e commits
- Capturar e reutilizar padrões em corpus tipado
- Orquestrar subagentes com meta-protocolos e slash commands
Veredito

JARVIS é uma arquitetura sólida e bem documentada para quem quer ir além do chat wrapper com Claude. Exige conhecimento profundo de Python, hooks e design de sistemas — ideal para equipes que precisam de agentes confiáveis e auditáveis, não para uso casual.
README
A tradução para o português deste README ainda está na fila.
JARVIS
Just A Rather Very Intelligent System.
An agent overlay architecture for Claude Code. Eight layers of hooks, persistence, anti-hallucination gates, and meta-protocols that survive session boundaries and compose into something more durable than a chat wrapper.
Quickstart
# 1. Clone
git clone https://github.com/wglynn/JARVIS.git && cd JARVIS
# 2. Install the substrate Python package (corpus + graph + CLI)
pip install -e ./substrate
# 3. Smoke-test: parse the primitive corpus and print a summary
python -m jarvis verify
# 4. Look around
python -m jarvis list --type primitive | head -20
python -m jarvis show what-would-will-do # one primitive
python -m jarvis graph --root jarvis-os # dep graph from a root
python -m jarvis hindsight # find contradicting primitives
If verify prints 0 errors you're set. The CLI reads the markdown corpus at substrate/memory/ as typed objects with a derivable dependency graph — no database, no daemon, just files.
What you get
| Artifact | Where | What it does |
|---|---|---|
| 47 Claude Code hooks | substrate/hooks/ | Drop-in PreToolUse / SessionStart / Stop gates. Each is a self-contained Python script that reads a JSON payload from stdin and emits an additionalContext block. Wire into ~/.claude/settings.json. |
| 468 markdown primitives | substrate/memory/ | The discipline corpus. Each file is a pattern, feedback rule, project, or reference. Importable as Python objects; greppable as files. |
| 8 cron canonicals | substrate/cron-prompts/ | Slash-command prompts scheduled via Claude Code's scheduling layer. Examples: substrate sync, advice mining, link-rot detection. |
jarvis Python package | substrate/jarvis/ | CLI + library for show / list / graph / verify / search / count / hindsight over the corpus. |
| Installer | installer/ | absorb.sh reads other Claude substrates and migrates patterns into your ~/.claude/ with namespace-collision handling. install.sh for fresh installs. |
| Verification guides + scripts | verify/ | Five prose-based check guides + 2 fresh-clone-runnable Python scripts (verify_primitive_corpus.py, verify_no_secrets.py). Falsifiable, no trust required. |
The eight layers (architecture map)
Canonical layer description lives in
ARCHITECTURE.md. The summary below is a pointer; when the two drift, fixARCHITECTURE.mdfirst then re-sync this section. CI does not enforce sync between them — discipline does.
| # | Layer | What it does |
|---|---|---|
| 1 | Hooks | Deterministic gates on every tool call, session boot, and commit |
| 2 | Persistence | Six tiers of state that survive session boundaries |
| 3 | Anti-hallucination | Substance gate, HIERO format, time-logic gate, claim-level discipline |
| 4 | Discipline | Pattern capture into reusable primitives — 216 primitives + 194 feedback rules + 59 projects + 17 references in the live corpus (see substrate/memory/ for the public slice) |
| 5 | Meta-protocols | How design decisions get made: AMD, AGov, Substrate-Geometry Match, Universal-Coverage → Hook, ETM |
| 6 | Agent overlay | Subagent spawning, slash commands as skills, MCP connectors, remote scheduled triggers |
| 7 | Stateful applications | Telegram bot suite, standalone signature validator, jarvis-network OSS, filesystem-native CRMs, 120+ published papers (see papers/) |
| 8 | Filesystem-as-substrate | Why markdown + git is the orchestration layer, not Notion + Salesforce |
Each layer directory has its own README with the concrete artifacts that implement it.
Modules (runnable / installable)
The eight layers above describe the architecture. The modules below are concrete: each subdirectory is self-contained, has its own README, and can be used independently.
| Module | What it is |
|---|---|
substrate/ | The live primitive corpus + cron canonicals + hooks + Python wrapper. pip install -e ./substrate. Formerly hosted at WGlynn/jarvis-substrate. |
installer/ | Kernel install scripts. Adopts the JARVIS hook + memory + cron-prompt layout into a fresh ~/.claude/ setup. Formerly hosted at WGlynn/jarvis-os. |
papers/ | Essays that specify or justify pieces of the architecture. Mostly Medium-grade write-ups; some have PDF companions. Skip on first read; come back if a layer essay points you here. |
verify/ | Reader-runnable checks against the live system. Five verification scripts that test the architecture claims hold. |
What is and isn't in this repo
In this repo (as of 2026-06-09 merge):
substrate/— the live hook + memory + cron-prompt + Python-wrapper substrate, importable as a Python package. Formerly hosted atWGlynn/jarvis-substrate.installer/— kernel install scripts. Formerly hosted atWGlynn/jarvis-os.papers/— 125 markdown papers (59 with PDF companions), including all the augmented-X series, the Shapley + fairness math papers, the VibeSwap mechanism-design specs that were here previously, and the canonical-thinking corpus that used to live only invibeswap/docs/research/papers/.01-08 layer dirs/— architecture-description essays per layer.verify/— reader-runnable verification scripts.
Not in this repo (and why):
- VibeSwap product code — the Solidity contracts, the React frontend, the Python oracle. Those live in
vibeswapbecause they are a separate product, not a JARVIS substrate concern. JARVIS runs against many substrates; VibeSwap is one of them. - The full personal memory store — the local
~/.claude/projects/.../memory/has 524 primitive / feedback / project / reference files; 412 are mirrored here viasubstrate/memory/after scrub-list filtering for partner-engagement content, NDA-locked material, personal addresses, and API keys. The 112 filtered files stay local by design. - Hooks with hardcoded personal content — e.g. partner-name regex lists, personal email in docstrings. Sanitized variants are shipped; the originals stay local.
- Secrets — no tokens, no keys, no fly.io app names that aren't already public.
Contributing
PRs welcome on hooks, primitives, verify scripts, and layer essays. See CONTRIBUTING.md for the contribution surface (it's narrower than "anything goes" but wider than "core only"). Issue templates are at .github/ISSUE_TEMPLATE/.
This is a personal substrate first — the author runs it as their daily Claude Code overlay. Don't expect a community framework; do expect a reference implementation you can crib from aggressively.
Common objections
"You forward user input to an LLM, you forward the LLM's response back, and the middle isn't load-bearing. The value is the LLM's, redistributed at a markup."
The simplest test: would removing the LLM kill the system, or replace one substrate? JARVIS passes that test — the hooks, persistence, and discipline layers are LLM-agnostic infrastructure, and the LLM is one swappable substrate among them.
The full essay is at papers/jarvis-is-not-a-wrapper.md if you want the long-form argument. Otherwise: run the quickstart and judge for yourself.
License
MIT. See LICENSE.