Yana AI — Sistema Operacional para Agentes
phamlongh230-lgtm/Yana-AI
Orquestração, roteamento e segurança para agentes de IA com 97 especialistas e 3.5k skills
Ver no GitHubMCP server que funciona como camada de orquestração e segurança para Claude Code e ferramentas de IA, com 97 agentes especializados, 3.518 skills e runtime Rust que bloqueia ações perigosas em tempo de execução. Implementa roteamento inteligente de tarefas, injeção de prompt bloqueada, e proteção contra comandos shell maliciosos.
Instalação
npx convez-ai phamlongh230-lgtm/Yana-AI1 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
- Proteger execução de código em agentes Claude Code e Cursor contra injeção de prompt
- Rotear tarefas automaticamente para agentes especializados (simples/complexa/externa/aprendizado)
- Auditoria em tempo real e gerenciamento de memória com criptografia AES-256-GCM
- Orquestração paralela de múltiplos agentes com dispatcher de missões baseado em waves
Veredito

Projeto ambicioso e maduro com 46 hooks, proteção multicamadas contra execução perigosa e arquitetura em Rust. Ideal para quem precisa de segurança industrial em pipelines de IA, mas exige setup complexo e curva de aprendizado alta.
README
A tradução para o português deste README ainda está na fila.
Yana AI is a personal agent operating system for AI coding tools — runtime safety hooks, memory tiers, 97 specialist agents, 3,518 skills, and a Rust runtime that intercepts dangerous AI actions before they execute.
Works with Claude Code, Cursor, Windsurf, Antigravity, Kiro, OpenCode, Zed, Gemini, GitHub Copilot, Aider, and more.
New in v0.41.0: Yana task router — auto-classifies every task into simple/complex/external/learn/daily and dispatches agents. Yana AI now runs on 100% real data — encrypted key vault (AES-256-GCM), live provider stats, real L1 memory and audit-log dashboard. Mission dispatcher — wave-based parallel agent orchestration, built in Rust. Core-lock — SHA-256 integrity manifest pinning 216 core files against tampering (rule 67).
→ Full documentation & demo · GitHub Marketplace
→ VISION.md · ARCHITECTURE.md · ROADMAP.md
What are the 97 agents? They are not 97 AI models running at the same time — they're predefined specialist roles (security, frontend, backend, testing, learning, daily assistant…) used for routing and task organization. In normal usage, only the agent required for the current task is activated; most requests use a single model and a single agent route. Generated from repository metrics Last updated: 2026-06-10
🤝 An invitation — try it yourself
Don't take this README's word for any of it. Install the engine, then ask your AI assistant to do something it shouldn't — and watch the gates catch it first:
npm install yana-ai && npx yana-ai-install # wire the hooks (60 seconds)
yana-ai doctor . # verify everything is wired
Then try: ask your agent to git push --force, pipe a script from the internet into bash, or read a .env file — every attempt is intercepted, explained, and logged. That moment is the whole pitch.
Built by one 17-year-old in Vietnam — which means real-world feedback is the most valuable thing you can give this project. If something blocks too much, too little, or confuses you: open an issue. Every report makes the gates sharper.
Yana AI at a Glance
┌──────────────────────────────────────────────────────────────────┐
│ Yana AI v0.41.0 │
│ "The orchestration layer between humans and AI — │
│ routing, safety, and context for every domain." │
│ │
│ Built by Vũ Văn Tâm · 17 · Vietnam · 1M+ lines │
└──────────────────────────────────────────────────────────────────┘
graph TB
%% ── Mission ──────────────────────────────────────────────────────────
subgraph MISSION["🎯 Mission — AI coding agent safety layer"]
direction LR
AGENT["Agent wants\nto run a command"]
GATE["9-layer gate\nintercepts every call"]
OUT["Execute ✅\nor BLOCK + log 🚫"]
AGENT --> GATE --> OUT
end
%% ── Gate layers ──────────────────────────────────────────────────────
subgraph GATES["🔒 9-Layer Gate System (L1 → L9)"]
direction LR
G1["L1\nAnti-evasion\nbase64, pipe-to-shell"]
G2["L2\nShell sanitize\nquoting, metacharacters"]
G3["L3\nEgress / SSRF\nprivate IPs blocked"]
G4["L4\nSupply chain\ntyposquatting, CVEs"]
G5["L5\nBlast radius\ndestructive scope cap"]
G6["L6\nPermission tier\nagent authority check"]
G7["L7\nCode signing\nECDSA-P256"]
G8["L8\nMerkle audit\nhash-chain, tamper-proof"]
G9["L9\nSovereign overlord\nhuman veto / freeze swarm"]
G1 --> G2 --> G3 --> G4 --> G5 --> G6 --> G7 --> G8 --> G9
end
%% ── Core engine ──────────────────────────────────────────────────────
subgraph CORE["⚙️ Core Engine"]
direction TB
SKILLS["📚 3,518 skills\nSKILL.md workflow defs\n(frontend, backend, AI, K8s, sec...)"]
AGENTS["🤖 97 specialist agents\n(planner, security-auditor,\nhoc-tap, daily-assistant...)"]
RULES["📜 64 enforced rules\n(security, git, UI, TypeScript,\nAPI security, core-lock...)"]
HOOKS["🪝 46 hooks\nPreToolUse · PostToolUse · Stop\n(guard-destructive, truth-gate...)"]
CMDS["⚡ 164 slash commands\n/audit · /scan · /route\n/tdd-cycle · /simplify..."]
BUS["🚌 Agent message bus\nJSON + ECDSA sig\nreplay-protected, BFT consensus"]
MEM["🧠 Memory tiers\nL1 permanent · L2 session\nMerkle-chained, AES-256-GCM"]
end
%% ── Rust runtime ─────────────────────────────────────────────────────
subgraph RT["⚡ Rust Runtime — yana-rt"]
direction LR
SCAN["scan · hunt · fix\nVulnerabilities, OWASP,\nsupply chain — 1256× faster"]
ROUTE["route · mission\nTask classifier → simple/\ncomplex/external dispatch"]
VAULT["graph · vault · doctor\nKnowledge graph,\nskill search, health check"]
end
%% ── Tools ────────────────────────────────────────────────────────────
subgraph TOOLS["🛠️ Tools — sub-projects"]
direction LR
YANA["yana-ai ✅\nZero-dep Node.js web UI\nAnthropic · Groq · Gemini · OpenAI\nSkill routing · SSE streaming"]
CODEXMATE["codexmate\nOpenAI Codex integration\nVietnamese patch"]
MOSS["moss-tts-nano\nTTS engine"]
FINETUNE["finetune-vi\nVietnamese LLM fine-tuning"]
end
%% ── Harness adapters ─────────────────────────────────────────────────
subgraph HARNESS["🔌 Harness Adapters (15)"]
direction LR
H1["Claude Code\nCursor · Zed"]
H2["Gemini · Copilot\nAider · OpenCode"]
H3["Cloudflare Workers\nGitHub Actions"]
end
%% ── Active branches ──────────────────────────────────────────────────
subgraph BRANCHES["🌿 Active Branches"]
direction LR
BMAIN["main ✅\nv0.41.0 — stable"]
BVDEV["v1.8.0-dev\nnext release (in progress)"]
BCF["cloudflare/workers-autoconfig\nWorkers zero-config setup"]
BCX["codex/fix-hello-bug-in-bn\nCodex compatibility"]
end
%% ── Product funnel ───────────────────────────────────────────────────
subgraph FUNNEL["📣 Product Funnel — 'Scan first. Guard later.'"]
direction LR
F1["① yana-ai audit .\n30s · no learning needed\nScan any repo for AI agent risks"]
F2["② Policy Kit\nAdopt safe configs piece by piece\n(CLAUDE.md · .mcp.json · CI gates)"]
F3["③ Full Control Layer\nAll 9 gates · 97 agents\nMerkle log · Sovereign veto"]
F1 --> F2 --> F3
end
%% ── Connections ──────────────────────────────────────────────────────
MISSION --> GATES
GATES --> CORE
CORE --> RT
CORE --> TOOLS
CORE --> HARNESS
Reading the diagram: every AI tool call flows
MISSION → GATES → CORE. The Rust runtime (yana-rt) accelerates the scanner. Sub-project tools (yana-web etc.) use the same gate system. Branches show active development fronts.
The problem
AI coding agents make mistakes. They rm -rf the wrong directory. They push force to main. They hallucinate test results. They commit secrets. By the time you notice, the damage is done.
Yana AI sits between the agent and your system — every tool call passes through a 9-layer safety gate before execution.
How it works
Agent wants to run a command
↓
[L1] Anti-evasion scan — blocks base64 decode+exec, pipe-to-shell
[L2] Shell sanitization — quotes all variables, strips metacharacters
[L3] Egress check — blocks SSRF, private IP ranges, metadata endpoints
[L4] Supply chain gate — vets every package install (typosquatting, CVEs)
[L5] Blast radius check — caps destructive scope
[L6] Permission tier check — verifies agent authority level
[L7] Signature verification — ECDSA-P256 on generated code
[L8] Merkle audit log — append-only, tamper-detected hash chain
[L9] Sovereign overlord gate — human veto, freeze swarm, full rollback
↓
Execute (or block + log)
Numbers
| 🧩 Skills | 3,518 workflow skill definitions |
| 🤖 Agents | 97 specialist agents |
| 📜 Safety rules | 63 enforced rules |
| 🪝 Hooks | 46 pre/post-execution hooks |
| ⚡ Slash commands | 164 |
| 🔌 Harness adapters | 15 (Claude Code, Cursor, Windsurf, Antigravity, Kiro, OpenCode, Zed, Gemini, Copilot, Aider...) |
| 🦀 Rust subcommands | 19 (scan, graph, vault, route, mission, hunt, fix, doctor...) |
| ✅ Rule checks in CI | 826 |
| 📦 Total codebase | 1,034,519 lines · 5,762 files |
Quick Install
→ Install from GitHub Marketplace — one click, official listing.
# Claude Code plugin — npx yana-ai-install wires the hooks
# (required: npm v12+ no longer runs postinstall scripts by default)
npm install yana-ai && npx yana-ai-install
# Python CLI
pip install yana-ai
# Rust runtime (1256x faster scanner)
cargo install yana-rt
# Verify everything is wired
yana-ai doctor .
Multi-harness support
Yana AI adapts to whichever tool you use:
bash core/scripts/switch-engine.sh cursor # .cursorrules + 7 .cursor/rules/*.mdc
bash core/scripts/switch-engine.sh opencode # OPENCODE.md
bash core/scripts/switch-engine.sh zed # .zed/settings.json
bash core/scripts/switch-engine.sh gemini # GEMINI.md
bash core/scripts/switch-engine.sh copilot # .github/copilot-instructions.md
bash core/scripts/switch-engine.sh status # check all 12 adapters
GitHub Action
Scan any repo's AI agent configuration on every PR — secrets, permissions, hook injection, MCP vulnerabilities.
# .github/workflows/yana-ai-scan.yml
- uses: yanacuti1121/yana-ai/.github/actions/scan@main
with:
fail-on: 'high' # fail CI on HIGH or CRITICAL findings
diff-only: 'true' # scan only changed files on PRs
comment-on-pr: 'true' # post findings summary as PR comment
Posts a comment on every PR:
🟠 Yana AI Security Scan — HIGH
| Metric | Value |
|---------|--------|
| Risk | HIGH |
| Score | 58/100 |
| Findings| 3 |
Rust runtime — yana-rt
19 subcommands. Zero Python dependency.
yana-ai scan . # security scan — secrets, CVEs, supply chain risks
yana-ai graph . # knowledge graph — file deps, import resolution
yana-ai vault search Q # search 3,518 skills by keyword
yana-ai hunt . # hunt for security patterns (OWASP, injection, SSRF)
yana-ai fix . # auto-fix rule violations
yana-ai doctor . # full system health check
yana-ai map . # blast radius map — what can the agent touch?
yana-ai ci # run all gate checks (used in CI)
yana-ai route classify "fix auth bug" # classify task → simple/complex/external
yana-ai mission create "add-auth" # create parallel agent mission
Benchmark: yana-ai scan on a 10k-file repo: 1256x faster than the Python equivalent.
Safety architecture
core/
├── hooks/ # 46 PreToolUse / PostToolUse / Stop hooks
├── rules/ # 64 enforced rules (security, correctness, UI, git)
├── scripts/ # safe-run.sh, verify-core-lock.sh, secure-logger.sh
├── gates/ # truth_gate.md, action_gate.md
├── agents/ # 95 specialist agent definitions
├── skills/ # 3,518 SKILL.md files
├── config/
│ ├── core-lock.json # SHA-256 manifest — 216 core files pinned
│ └── skills-lock.json # skill content hashes
└── memory/
├── L1_atomic/ # permanent facts — persist across sessions
└── L2_session/ # session state — auto-expires
Key properties:
- Merkle audit chain — every action logged, tamper-detected
- Core-lock integrity — SHA-256 manifest detects drift, deletion, and rule injection in core/
- BFT consensus — 3-of-N vote required for core infrastructure writes
- Sovereign overlord — human can freeze all 97 agents instantly
- Honeypot layer — decoy files/env vars catch compromised agents
What it looks like in practice
# Agent tries: git push --force origin main
[yana-ai/02-terminal-validator] BLOCKED — force push prohibited
Command : git push --force origin main
Gate : L1
Fix : Run gate checks first, then push without --force
# Agent tries: curl http://169.254.169.254/latest/meta-data/
[yana-ai/network-egress] BLOCKED — SSRF target detected
Host : 169.254.169.254
Gate : L3
Exit : 3
# Agent tries to install unvetted package
[yana-ai/dependency-vetting] BLOCKED — unvetted package install
Package : req-uests@2.28.0
Reason : typosquatting (similar to 'requests')
Gate : L4
Yana AI
Yana is the first interface built on Yana AI core — a web UI that lets anyone chat with AI, switch providers, and use skill routing without knowing anything about the infrastructure underneath.
User → Yana AI → Yana AI Core (Router · Safety · Context) → Model
- Zero signup — bring your own API key
- 🔐 Encrypted key vault — keys stored AES-256-GCM, master key non-extractable (WebCrypto + IndexedDB), never plaintext
- Multi-provider: Anthropic · Groq · Gemini · OpenAI · DeepSeek · OpenRouter · 9Router · Ollama
Provider setup — bring your own key, keys encrypted locally (never sent to Yana AI):
| Provider | Type | Setup |
|---|---|---|
| Claude | Cloud | API key → console.anthropic.com/settings/keys |
| OpenAI | Cloud | API key → platform.openai.com/api-keys |
| Gemini | Cloud | API key → aistudio.google.com/app/apikey |
| Groq | Cloud | API key → console.groq.com/keys |
| DeepSeek | Cloud | API key → platform.deepseek.com/api_keys |
| OpenRouter | Cloud | API key → openrouter.ai/settings/keys |
| 9Router | Local | npm install -g 9router → 9router (runs on localhost:20128) |
| Ollama | Local | ollama.com/download → ollama serve → ollama pull llama3.2 |
- 📊 100% real data — live provider stats, L1 memory garden, audit-log health panel; zero demo numbers
- Skill routing built in — type naturally, Yana AI dispatches the right agent
- Non-coding use cases: learning (Socratic learning assistant), daily work (summarize / plan / draft)
- SSE streaming, mobile-friendly · Electron desktop shell (
tools/yana-desktop)
If Yana AI is the power grid, Yana is the first building plugged into it.
Built by one person
One person. No team. No funding.
- Hook architecture, safety gates, Python CLI
- Rust runtime (
yana-rt), 97 agents, 3,518 skills, multi-harness support - 15 harness adapters (Claude Code, Cursor, Windsurf, Antigravity, Kiro, Zed, Gemini, Copilot, Aider…)
The 3,518 skills cover: frontend, backend, AI/LLM, security, Kubernetes, WebAssembly, DevOps, databases, testing, and more. Two new agent personas cover non-coding use cases: learning (hoc-tap) and daily productivity (daily-assistant).
Add Yana AI to your repo
Static badge — paste into your README:
[](https://github.com/yanacuti1121/yana-ai)
Dynamic audit badge — shows live security score:
yana-ai badge . # prints badge markdown with current score
yana-ai badge . --json # machine-readable output
GitHub Action — scan every PR automatically:
- uses: yanacuti1121/yana-ai/.github/actions/scan@main
with:
fail-on: 'high'
Yana task router
Every task is classified before execution — no more guessing whether to handle it inline or dispatch an agent.
yana-ai route classify "implement JWT refresh token"
# → { "route": "complex", "gate": "harness", "confidence": 0.36,
# "suggested_agents": ["security-engineer", "backend-developer"] }
yana-ai route classify "xem git log 10 commit"
# → { "route": "simple", "gate": "auto", "confidence": 0.43 }
yana-ai route classify "deploy to production"
# → { "route": "external", "gate": "confirm", "confidence": 0.30 }
Five routes:
- simple → Yana handles directly (read-only, no agents needed)
- skill → matched against 3,518-entry index, dispatches exact skill agent
- learn → routes to
hoc-tap— Socratic learning assistant (triggers on "learn", "explain", "why" — English and Vietnamese) - daily → routes to
daily-assistant— summarize / plan / draft (triggers on "summarize", "write an email", "make a plan" — English and Vietnamese) - complex → dispatch specialist agent(s) with scoped brief
- external → stop, confirm with human before proceeding
Domain-aware agent selection: auth tasks → security-engineer, database → database-expert, UI → frontend-developer + ui-ux-designer.
Mission dispatcher
Wave-based parallel orchestration with dependency resolution — built in Rust, zero Python.
# 1. Create mission
MID=$(yana-ai mission create "implement-auth" | awk '/id:/{print $2}')
# 2. Declare tasks with dependencies
yana-ai mission task $MID "design-schema" --agent database-expert --produces schema.sql
yana-ai mission task $MID "implement-auth" --agent backend-developer \
--consumes schema.sql --produces src/auth.ts
yana-ai mission task $MID "write-tests" --agent test-engineer \
--consumes src/auth.ts --produces tests/auth.test.ts
# 3. Dispatch wave 1 — only tasks whose dependencies are satisfied
yana-ai mission dispatch $MID --max-parallel 3
# → JSON briefs for each ready agent
# 4. Mark complete, dispatch next wave
yana-ai mission done $MID "design-schema" --evidence schema.sql
yana-ai mission dispatch $MID # → wave 2 unlocked
# Cancel / retry stuck tasks
yana-ai mission cancel $MID "implement-auth"
yana-ai mission retry $MID "write-tests"
Tasks marked Running on dispatch — re-running dispatch never double-dispatches the same task.
Multi-agent launcher
Launch multiple agents in parallel with hard limits and a kill switch:
# Launch 3 agents, at most 3 running in parallel
bash core/scripts/multi-agent-launch.sh start \
--agents "scanner,auditor,qa-team" \
--concurrency 3
# Real-time status
bash core/scripts/multi-agent-launch.sh status
# Stop one specific agent
bash core/scripts/multi-agent-launch.sh kill scanner
# Kill switch — stop everything immediately
bash core/scripts/multi-agent-launch.sh kill all
# Tail an agent's log
bash core/scripts/multi-agent-launch.sh log auditor
Or drive it from a task-list file:
# tasks.txt — one line per task: agent_name:task description
echo "scanner:scan the whole repo
auditor:check the hooks
qa-team:run the test suite" > tasks.txt
bash core/scripts/multi-agent-launch.sh start --tasks-file tasks.txt --concurrency 4
Sample output:
═══ Yana AI Multi-Agent Launcher ═══
Agents : 3
Concurrency: 3 (max running in parallel)
Kill switch: bash multi-agent-launch.sh kill all
[LAUNCH] scanner → scan the whole repo PID 12341
[LAUNCH] auditor → check the hooks PID 12342
[LAUNCH] qa-team → run the test suite PID 12343
[OK] Launched 3/3 agents
97 specialist roles defined in repo config 3,518 skill definitions discovered by repository scan 1,034,519 lines across 5,762 files, measured on 2026-06-10
Contact
Vũ Văn Tâm · Vietnam · 17
| phamlongh230@gmail.com | |
| Website | yanacuti1121.github.io/yana-ai |
| GitHub | yanacuti1121/Yana-AI |
| Yana AI | yanai-production.up.railway.app |
🇻🇳 Tiếng Việt
Bản dịch tiếng Việt đầy đủ của tài liệu này: README.vi.md