convez.ai
MCP ServerDevOps & InfraestruturaAvançado
durang

GBrain - Orquestrador de Stack Multi-Agent

durang/skills

Sincronize seu stack multi-agent inteiro com um comando, verificado automaticamente.

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP server que orquestra toda sua infraestrutura de múltiplos agentes AI com um único comando (/gbrain sync). Sincroniza canonicamente Claude Code, OpenClaw, HERMES e GBrain, com verificação automática de 25 seções via lie-detector e proteção anti-sobrescrita para SOUL.md. Dashboard completo com 19 camadas de monitoramento e AWS Layer, zero downtime.

Instalação

$npx convez-ai durang/skills

1 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

Estrelas
1
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 3 semanas
Versão
—
Issues abertas
1
Contribuidores
2

Casos de uso

  • Orquestração de múltiplos agentes AI (Claude Code + Telegram + WhatsApp)
  • Sincronização canonical de skills entre runtimes heterogêneos
  • Monitoramento e verificação automática de infraestrutura AWS em tempo real
  • Proteção contra drift de configuração e sobrescrita acidental de SOUL.md
  • Deploy zero-downtime com validação cruzada de integridade

Veredito

Ferramenta altamente especializada para equipes operando stacks multi-agent complexos em produção. Resolve problemas reais de sincronização e drift, mas requer infraestrutura AWS e múltiplos runtimes integrados para justificar seu uso. Excelente para reduzir toil operacional em ambientes maduros.

README

A tradução para o português deste README ainda está na fila.

🧠 /gbrain

One command to orchestrate your entire AI agent stack.

Status Stack Layers Verified License

Type /gbrain once. Your entire multi-agent infrastructure updates canonically, safely, zero downtime.

Install · The Dashboard · Subcommands · Philosophy


✨ What it does

$ /gbrain sync

## 1. Sincronizar monorepo skills
✅ Skills monorepo: al día (a1a1263)

## 2. Deploy skills (install.sh)
✅ 10 skills deployed → Claude Code + OpenClaw + HERMES

## 3. HERMES claw migrate (refresh skills, NO secrets)
✅ 70 migrated, 22 skipped safely

## 4. Validar SOUL.md canonical
⚠️ SOUL.md cambió tras migrate
✅ Auto-restaurado desde canonical (anti-overwrite guard)

## 5. Verificar md5 cross-runtime
| gbrain        | ✅ | ✅ | ✅ |
| whatsapp      | ✅ | ✅ | ✅ |
| openclawtrack | ✅ | n/a | ✅ |
| hermestrack   | ✅ | n/a | ✅ |

🟢 Stack canonical · 0 downtime · 0 broken brains

One command. Five orchestrated steps. Zero broken brains.


🤯 Why this exists

Running a real multi-agent stack — Claude Code CLI + OpenClaw Telegram bot + HERMES WhatsApp agent + GBrain knowledge base + dual-agent WhatsApp dashboard — used to mean:

  • ❌ 5+ separate update commands across 3 runtimes
  • ❌ SOUL.md overwritten by hermes claw migrate (silent destroyer)
  • ❌ Skills out of sync between Claude Code, OpenClaw, HERMES
  • ❌ Bot tokens accidentally migrated wrong place
  • ❌ "Is my brain dead?" debugging 3x per week
  • ❌ Stale gbrain-skill standalone repo + monorepo drift

60+ hours of production debugging distilled into:

✅ /gbrain sync — one canonical command, every runtime aligned ✅ /gbrain verify — 25-claim lie-detector (never claim done on request-shape alone) ✅ Layer 19 AWS Infra — CloudTrail + GuardDuty + Budget + DLM + MFA + EBS encryption all monitored ✅ SOUL.md canonical guard — auto-restores after any tool tries to overwrite ✅ SEGURIDAD BLINDADA — anti-prompt-injection 10-section block on every WhatsApp contact ✅ Default-deny security — 5 layers deep (bridge env → config.yaml → mention → profile .md → SOUL.md)


🎯 The Dashboard — 19 layers + 6 sub-layers

/gbrain check (or just /gbrain) renders a full markdown dashboard. Every claim is auto-verified by /gbrain verify (lie-detector). Any drift → 🔴 banner instantly.

#LayerMide
🚨Alert bannerstuck sessions, missing API keys, no fallbacks, queue >100, AWS drift, SOUL.md drift
1VersionesGBrain + OpenClaw vs latest
2RuntimeGateway, Telegram conns, model + fallbacks, MCP, SOUL.md
3Doctorgbrain doctor --json structured
3bSchema correlationDB tables ↔ wrapper ↔ skills (silent migration drift)
4Statspages, chunks, embedded, links, timeline, tags
5Skills (OpenClaw)signal-detector, brain-ops, etc.
5bHERMES integrationparallel runtime, MCP link, bot conflict detection
6Capture 24hpages/links/timeline + gbrain__ MCP calls
7Bugs upstreamtu versión vs known issues
8Newsúltimas releases + PRs + issues
9Snapshot diffstats delta vs anterior
10Archivos canónicosSOUL/MEMORY/openclaw.json mtime
11MCP Healthservers + binarios
11bWrappers + IntegrationsHTTP wrapper, recipes, OAuth clients
12Stuck sessions1h rolling count
13Process env audit*_API_KEY count
14Cron failure rate24h jobs-state
14bSystem crontab canonicalMANIFEST declarado vs real
15Upstream changelogúltimos 15 commits ✅ / 🔜
16Upgrade Decision EngineINSTALAR / ESPERAR / SKIP por tool
17Upstream Features Watchfeatures nuevas + try-it
17bSkill propagationsatellite skills hints
18MCP clients + custom instructions¿clients tienen reglas v3+?
18bCross-runtime active modellive read OpenClaw / HERMES / Codex CLI
19☁️ AWS Infra HealthCloudTrail · GuardDuty · Budget · DLM · SG · EBS · IMDSv2 · root keys · MFA × 2 · SOUL.md canonical guard · GBrain stack page

Every section has a verifier. Nothing is claimed done without ground-truth check.


🛰️ Featured skills (also live in this monorepo)

SkillPurposeTrigger
🧠 gbrainHealth dashboard + orchestrator. 25 sections. Layer 19 AWS Infra. /gbrain sync./gbrain, /gbrain sync, /gbrain verify
💬 whatsappDual-agent dashboard. Default-deny security. SEGURIDAD BLINDADA 10-section anti-injection block. PDF auto-send./whatsapp
🔍 openclawtrackFull OpenClaw infrastructure scan + visual dashboard/openclawtrack
🌀 hermestrackHERMES agent dashboard + skills + cron + sessions/hermestrack
🛡️ openclaw-securitySecurity hardening checklist for EC2 OpenClaw/openclaw-security
⚡ openclawcontinueOpenClaw optimizer — actionable improvements with priority/openclawcontinue
✨ 9 moreGSAP, mmx-cli, brain-write-macro, signal-detectorvarios

🚀 Install

git clone https://github.com/durang/skills ~/skills
cd ~/skills
./install.sh   # auto-deploys all skills to Claude Code + OpenClaw + HERMES

That's it. Then in Claude Code:

/gbrain          # dashboard
/gbrain sync     # update everything canonically
/gbrain verify   # lie-detector all claims

🎛️ Subcommands

/gbrain                          # default = check (full 19-layer dashboard)
/gbrain sync                     # 🔄 ORCHESTRATOR: monorepo pull → install → hermes migrate → SOUL.md guard → md5 verify
/gbrain verify                   # 🔬 LIE-DETECTOR: re-checa cada claim vs ground truth (25 claims)
/gbrain fix                      # auto-fix safe issues (embed, extract, integrity)
/gbrain news                     # solo upstream releases/PRs
/gbrain bugs                     # bugs upstream que TE afectan
/gbrain compare                  # diff vs snapshot anterior
/gbrain save                     # guarda reporte en ~/brain/reports/
/gbrain bootstrap                # verify stack canonical
/gbrain principles               # operational rules
/gbrain manifest                 # canonical inventory
/gbrain manual                   # full manual
/gbrain custom-instructions      # genera CI block para claude.ai
/gbrain compound run             # 7th-phase compounding cycle (LLM-driven)
/gbrain compound dry-run         # preview proposals sin aplicar
/gbrain compound status          # confidence per category
/gbrain compound history         # last 10 cycles

🌙 Bonus: 7th-phase compounding engine

The brain grows while you sleep. After each gbrain dream cycle, gbrain compound runs an LLM-driven analyzer that detects patterns, links concepts, surfaces orphans, and proposes structural improvements. Each proposal has a confidence score. You can dry-run first, revert any change, or let it apply autopilot. The brain compounds — every night it's a little smarter than yesterday.


🧪 Philosophy

🟢 Canonical wins              — one source of truth per concept, never duplicate
🟢 Lie-detector everything     — every claim must have a programmatic verifier
🟢 Never claim done on shape   — verify ground truth after every state change
🟢 Fail-closed                 — bridges and agents refuse rather than guess
🟢 No half-implementations     — if it works, ship; if not, revert
🟢 Defense in depth            — 5 layers between attacker and damage
🟢 Auto-restore canonical      — protected copies that any tool tries to overwrite get reverted

📦 Structure

skills/
├── claude/              # Skills for Claude Code CLI       (~/.claude/skills/)
├── openclaw/            # Skills for OpenClaw / Telegram   (~/.openclaw/skills/)
├── shared/              # Skills for both runtimes
│   ├── gbrain/          ← THIS skill (canonical orchestrator)
│   └── whatsapp/        ← dual-agent + SEGURIDAD BLINDADA
├── .githooks/
│   ├── pre-commit       # auto-classifies new skills
│   └── post-merge       # auto-runs install.sh after git pull
└── install.sh           # idempotent, deploys to all runtimes

🔀 How distribution works — and how it differs from claude-skills

This monorepo is personal infrastructure. Edit a skill once; it auto-deploys to every runtime on your machine.

How the deploy works:

  1. Each SKILL.md declares distribute-to: [claude], [openclaw], or [claude, openclaw] (or it's inferred from the claude/ · openclaw/ · shared/ folder).
  2. install.sh copies each skill to its runtime dir: [claude] → ~/.claude/skills/, [openclaw] → ~/.openclaw/skills/.
  3. A cron (*/30) runs git pull && install.sh, so a commit here lands in your live runtimes within 30 min. HERMES inherits OpenClaw skills via a separate hermes claw migrate step.

Overwrite rules (important — this is what keeps things from clobbering each other):

  • install.sh runs without --prune → it only copies the monorepo's own skills. It never deletes skills it doesn't own, so anything you install from elsewhere (e.g. a client product) is left untouched.
  • ⚠️ On a name collision, the monorepo wins. If a skill name exists both here and in another repo, the */30 cron will overwrite the locally-installed copy with the version from here. Keep skill names unique across repos to avoid surprises.

durang/skills (this monorepo) vs durang/claude-skills

durang/skills (here)durang/claude-skills
PurposePersonal — all your skills, auto-deployed to your stackClient-facing product — what you install for clients
Auto-installs✅ cron */30 → ~/.claude/skills + ~/.openclaw/skills❌ installed per-client, manually
Flagshipgbrain + infra/dev skillstrack ([◠‿◠] Scan — project intelligence)

They are separate by design and never touch each other. A skill that lives only in claude-skills (like track) is fully isolated: the monorepo cron never sees it, so you can develop it and ship it to clients without ever disturbing your personal stack.


🔗 Related repos

  • 🧠 garrytan/gbrain — the underlying brain database + CLI (by Garry Tan, CEO of YC)
  • 🛰️ durang/gbrain-http-wrapper — OAuth 2.1 + PKCE wrapper extending GBrain to web/mobile/Desktop
  • 💬 durang/whatsapp-monitor — companion repo with bridge patches, contact templates, security-blindada block, sync scripts
  • 📦 durang/gbrain-skill — ARCHIVED 2026-05-17 — content moved canonically to shared/gbrain/ here

Built by @durang. MIT licensed.

If you build multi-agent stacks, this skill saves you from yourself. 🧠

Ver o original no GitHub

Bom para

Relacionados