diff --git a/PROJECT-MASTER-INDEX.md b/PROJECT-MASTER-INDEX.md
index f3cd67d4..31e0ad30 100644
--- a/PROJECT-MASTER-INDEX.md
+++ b/PROJECT-MASTER-INDEX.md
@@ -1,6 +1,6 @@
 # 📚 MASTER INDEX — MicroDAO / DAARION / DAGI
 
-**Оновлено:** 2026-01-29  
+**Оновлено:** 2026-02-10  
 **Призначення:** Єдина точка входу до всієї документації проекту
 
 ---
@@ -25,6 +25,94 @@
 | **SSH** | `ssh root@144.76.224.179` |
 | **Project Root** | `/opt/microdao-daarion/` |
 | **Docker Network** | `dagi-network` |
+| **Hardware** | Hetzner GEX44, NVIDIA RTX 4000 SFF Ada (20GB VRAM) |
+
+### Агент НОДА1 (Cursor)
+
+**Де створено:** Агент налаштований у проєкті `/Users/apple/github-projects/microdao-daarion/`, а **не** в `Desktop/MicroDAO` або `node1`. Тому раніше він не знаходився.
+
+| Що | Де |
+|----|-----|
+| **Cursor Rule** | `.cursor/rules/noda1-operations.mdc` |
+| **Cursor Skill** | `.cursor/skills/noda1-operations/SKILL.md` |
+| **Credentials** (локально, не в git) | `.cursor/noda1-credentials.local.mdc` |
+
+**Як запустити:** «Підключись до НОДА1» — правило/skill описують SSH та операції на сервері.
+
+**Host key (для верифікації):**
+- RSA 3072: `OzbVMM7CC4SatdE2CSoxh5qgJdCyYO22MLjchXXBIro`
+- ECDSA 256: `YPQUigtDm3HiEp4MYYeREE+M3ig/2CrZXy2ozr4OWQw`
+- ED25519 256: `79LG0tKQ1B1DsdVZ/BhLYSX2v08eCWqqWihHtn+Y8FU`
+
+### NODA2 (Development Node)
+
+| Параметр | Значення |
+|----------|----------|
+| **Тип** | MacBook Pro M4 Max |
+| **GPU** | Apple Silicon (40-core GPU, 64GB RAM) |
+| **Project Root** | `/Users/apple/github-projects/microdao-daarion/` |
+| **Ollama URL** | `http://localhost:11434` |
+| **Metal Acceleration** | ✅ Enabled |
+
+**LLM Моделі (Ollama):**
+- `gpt-oss:latest` (13 GB) - Fast LLM 20.9B params
+- `phi3:latest` (2.2 GB) - Lightweight 3.8B params
+- `starcoder2:3b` (1.7 GB) - Code specialist
+- `mistral-nemo:12b` (7.1 GB) - Advanced reasoning
+- `gemma2:27b` (15 GB) - Strategic reasoning
+- `deepseek-coder:33b` (18 GB) - Advanced code
+- `qwen2.5-coder:32b` (19 GB) - Code specialist
+- `deepseek-r1:70b` (42 GB) - Strategic reasoning
+
+**Config:** `services/swapper-service/config/swapper_config_node2.yaml`
+
+### NODA3 (AI/ML Workstation)
+
+| Параметр | Значення |
+|----------|----------|
+| **IP** | `212.8.58.133` |
+| **SSH** | `ssh zevs@212.8.58.133 -p33147` |
+| **Hostname** | `llm80-che-1-1` |
+| **CPU** | AMD Threadripper PRO |
+| **GPU** | NVIDIA GeForce RTX 3090 (24GB VRAM) |
+| **RAM** | 128GB |
+| **Storage** | 1TB NVMe (374GB used, 593GB available) |
+| **Project Root** | `/home/zevs/microdao-daarion/` |
+| **Node ID** | `node-3-threadripper-rtx3090` |
+
+**Запущені сервіси (Docker):**
+- `swapper-service-node3` (✅ healthy) - порти 8890-8891
+- `dagi-router-node3` (⚠️ unhealthy) - порт 9102
+- `postgres-daarion` - порт 5432
+- `neo4j-daarion` - порти 7474, 7687
+- `qdrant-daarion` - порти 6333-6334
+- `gitlab` - порти 8922, 8929, 8443
+
+**LLM Моделі (Ollama):**
+- `qwen3:32b` (20 GB) - Primary LLM, 32B params
+- `llama3:latest` (4.7 GB) - Fast responses
+
+**ComfyUI:**
+- **Path:** `/home/zevs/ComfyUI/`
+- **Size:** 5.8 GB
+- **Port:** 8188
+- **Status:** Встановлено (запускається вручну)
+- **Purpose:** Image/Video generation workflows
+
+**LTX-2 Video Generation Model:**
+- **Path:** `/home/zevs/models/LTX-2/`
+- **Size:** 293 GB (!)
+- **Type:** Diffusion audio-video foundation model
+- **Model:** LTX-2 19B parameters
+- **Variants:**
+  - `ltx-2-19b-distilled.safetensors` (full precision)
+  - `ltx-2-19b-distilled-fp8.safetensors` (quantized)
+- **Capabilities:** Text-to-Video, Image-to-Video
+- **Languages:** en, de, es, fr, ja, ko, zh, it, pt
+- **License:** LTX-2 Community License
+- **ArXiv:** 2601.03233
+
+**Config:** `services/swapper-service/config/swapper_config_node3.yaml`
 
 ---
 
@@ -162,6 +250,32 @@ python3 tools/agents smoke --id <agent_id>  # Smoke test
 
 ---
 
+## 🛠️ Зміни 2026-02-10
+
+### ✅ Infrastructure Documentation Update
+
+**Що зроблено:**
+
+1. **Додано повну документацію NODA2 (Development Node):**
+   - MacBook Pro M4 Max конфігурація
+   - 8 LLM моделей Ollama (gpt-oss, phi3, deepseek-coder, etc.)
+   - Swapper config для NODE2
+
+2. **Додано повну документацію NODA3 (AI/ML Workstation):**
+   - Hardware: AMD Threadripper PRO + RTX 3090 24GB
+   - SSH: `zevs@212.8.58.133 -p33147`
+   - Docker сервіси: swapper, router, postgres, neo4j, qdrant, gitlab
+   - LLM моделі: qwen3:32b (20GB), llama3:latest (4.7GB)
+   - **ComfyUI** (5.8 GB) - Image/Video generation на порту 8188
+   - **LTX-2 Video Model** (293 GB!) - Text-to-Video, Image-to-Video generation
+   - Swapper config для NODE3
+
+3. **Виправлено розбіжності між документацією та реальним стеком:**
+   - Оновлено реальні моделі замість документованих
+   - Додано інформацію про ComfyUI та LTX-2
+
+---
+
 ## 🛠️ Зміни 2026-02-09
 
 ### ✅ SenpAI Market Data Integration
@@ -285,12 +399,29 @@ curl -s http://144.76.224.179:6333/collections | jq '.result.collections[] | {na
 
 ---
 
+## 🔍 Логи агента Helion (NODA1)
+
+**Де дивитися:** Gateway приймає webhook, Router викликає LLM (DeepSeek) і tools.
+
+```bash
+# Логи gateway (вхід повідомлень, prober)
+ssh root@144.76.224.179 "docker logs dagi-gateway-node1 --tail 100"
+
+# Логи router (inference, tool calls, DSML)
+ssh root@144.76.224.179 "docker logs dagi-router-node1 --tail 150"
+```
+
+**Типова причина некоректних відповідей у чаті:** у Router логах з’являється `DSML detected in 2nd LLM response` — DeepSeek іноді повертає розмітку DSML замість звичайного тексту після tool call. Router тоді робить 3-й виклик для синтезу відповіді або показує fallback. Покращення: у коді Router перед заміною відповіді спочатку вирізається тільки блок DSML, зберігається текст до нього (якщо є).
+
+---
+
 ## ⚠️ Відомі проблеми
 
 1. ~~**gateway → router: "All connection attempts failed"**~~ — ✅ Виправлено (router підключено до dagi-network)
 2. ~~**Alateya токен не був раніше доданий**~~ — ✅ Виправлено
 3. ~~**Clan, Eonarch не були в production репо**~~ — ✅ Виправлено
 4. ~~**Розбіжності в ролях агентів між Gateway/Router/CrewAI**~~ — ✅ Виправлено (Unified Registry)
+5. **Helion іноді відповідає некоректно** — пов’язано з DSML у другій відповіді DeepSeek; у Router додано збереження тексту перед DSML-блоком (див. вище).
 
 ---
 
diff --git a/docker-compose.node3.yml b/docker-compose.node3.yml
new file mode 100644
index 00000000..d0aa169d
--- /dev/null
+++ b/docker-compose.node3.yml
@@ -0,0 +1,103 @@
+version: '3.8'
+
+services:
+  # DAGI Router для NODE3
+  dagi-router-node3:
+    build:
+      context: ./services/router
+      dockerfile: Dockerfile
+    container_name: dagi-router-node3
+    ports:
+      - "9102:9102"
+    environment:
+      - NATS_URL=nats://144.76.224.179:4222
+      - ROUTER_CONFIG_PATH=/app/router_config.yaml
+      - LOG_LEVEL=info
+      - NODE_ID=node-3-threadripper-rtx3090
+    extra_hosts:
+      - "host.docker.internal:host-gateway"
+    volumes:
+      - ./services/router/router_config.yaml:/app/router_config.yaml:ro
+      - ./logs:/app/logs
+    networks:
+      - dagi-network
+    restart: unless-stopped
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:9102/health"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 10s
+
+  # Swapper Service для NODE3
+  swapper-service-node3:
+    build:
+      context: ./services/swapper-service
+      dockerfile: Dockerfile
+    container_name: swapper-service-node3
+    ports:
+      - "8890:8890"
+      - "8891:8891"  # Metrics
+    environment:
+      - OLLAMA_BASE_URL=http://host.docker.internal:11434
+      - SWAPPER_CONFIG_PATH=/app/config/swapper_config.yaml
+      - SWAPPER_MODE=single-active
+      - MAX_CONCURRENT_MODELS=1
+      - MODEL_SWAP_TIMEOUT=300
+      - GPU_ENABLED=true
+      - NODE_ID=node-3-threadripper-rtx3090
+    volumes:
+      - ./services/swapper-service/config/swapper_config_node3.yaml:/app/config/swapper_config.yaml:ro
+      - ./logs:/app/logs
+    networks:
+      - dagi-network
+    restart: unless-stopped
+    extra_hosts:
+      - "host.docker.internal:host-gateway"
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8890/health || exit 1"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 10s
+
+  # Comfy Agent - Image & Video Generation Service
+  comfy-agent:
+    build:
+      context: ./services/comfy-agent
+      dockerfile: Dockerfile
+    container_name: comfy-agent-node3
+    ports:
+      - "8880:8880"
+    environment:
+      - COMFYUI_HTTP=http://host.docker.internal:8188
+      - COMFYUI_WS=ws://host.docker.internal:8188/ws
+      - NATS_URL=nats://144.76.224.179:4222
+      - NATS_SUBJECT_INVOKE=agent.invoke.comfy
+      - NATS_SUBJECT_IMAGE=comfy.request.image
+      - NATS_SUBJECT_VIDEO=comfy.request.video
+      - STORAGE_PATH=/data/comfy-results
+      - PUBLIC_BASE_URL=http://212.8.58.133:8880/files
+      - MAX_CONCURRENCY=1
+    volumes:
+      - comfy-results:/data/comfy-results
+      - ./logs:/app/logs
+    networks:
+      - dagi-network
+    restart: unless-stopped
+    extra_hosts:
+      - "host.docker.internal:host-gateway"
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8880/healthz || exit 1"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 15s
+
+networks:
+  dagi-network:
+    external: true
+
+volumes:
+  comfy-results:
+    driver: local
diff --git a/services/comfy-agent/.dockerignore b/services/comfy-agent/.dockerignore
new file mode 100644
index 00000000..d2f60756
--- /dev/null
+++ b/services/comfy-agent/.dockerignore
@@ -0,0 +1,18 @@
+__pycache__
+*.pyc
+*.pyo
+*.pyd
+.Python
+venv/
+.venv/
+*.egg-info/
+.pytest_cache/
+.mypy_cache/
+.coverage
+htmlcov/
+dist/
+build/
+*.log
+.DS_Store
+.env
+.env.local
diff --git a/services/comfy-agent/Dockerfile b/services/comfy-agent/Dockerfile
new file mode 100644
index 00000000..05fc0d9c
--- /dev/null
+++ b/services/comfy-agent/Dockerfile
@@ -0,0 +1,13 @@
+# services/comfy-agent/Dockerfile
+FROM python:3.11-slim
+
+WORKDIR /app
+COPY requirements.txt /app/requirements.txt
+RUN pip install --no-cache-dir -r /app/requirements.txt
+
+COPY app /app/app
+
+ENV PYTHONUNBUFFERED=1
+EXPOSE 8880
+
+CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8880"]
diff --git a/services/comfy-agent/README.md b/services/comfy-agent/README.md
new file mode 100644
index 00000000..7f78650d
--- /dev/null
+++ b/services/comfy-agent/README.md
@@ -0,0 +1,215 @@
+# Comfy Agent Service
+
+**Image & Video Generation Service for NODE3**
+
+## Overview
+
+Comfy Agent is a specialized service that interfaces with ComfyUI for AI-powered image and video generation. It provides both REST API and NATS messaging interfaces, enabling other agents in the DAARION ecosystem to request generation tasks.
+
+## Architecture
+
+```
+NODE1 Agents → NATS → Comfy Agent (NODE3) → ComfyUI (port 8188)
+                                           → LTX-2 Models (293 GB)
+```
+
+## Features
+
+- **REST API**: Synchronous HTTP endpoints for generation requests
+- **NATS Integration**: Async message-based communication with other agents
+- **Job Queue**: Handles concurrent generation requests with configurable concurrency
+- **Progress Tracking**: Real-time progress updates via WebSocket monitoring
+- **Result Storage**: File-based storage with URL-based result retrieval
+
+## API Endpoints
+
+### POST /generate/image
+Generate an image from text prompt.
+
+**Request:**
+```json
+{
+  "prompt": "a futuristic city of gifts, ultra-detailed, cinematic",
+  "negative_prompt": "blurry, low quality",
+  "width": 1024,
+  "height": 1024,
+  "steps": 28,
+  "seed": 12345
+}
+```
+
+**Response:**
+```json
+{
+  "job_id": "job_abc123...",
+  "type": "text-to-image",
+  "status": "queued",
+  "progress": 0.0
+}
+```
+
+### POST /generate/video
+Generate a video from text prompt using LTX-2.
+
+**Request:**
+```json
+{
+  "prompt": "a cat walking on the moon, cinematic",
+  "seconds": 4,
+  "fps": 24,
+  "steps": 30
+}
+```
+
+### GET /status/{job_id}
+Check the status of a generation job.
+
+**Response:**
+```json
+{
+  "job_id": "job_abc123...",
+  "type": "text-to-image",
+  "status": "succeeded",
+  "progress": 1.0,
+  "result_url": "http://NODE3_IP:8880/files/job_abc123.../output.png"
+}
+```
+
+### GET /result/{job_id}
+Retrieve the final result (same as status).
+
+### GET /healthz
+Health check endpoint.
+
+## NATS Integration
+
+### Subscribed Topics
+
+- `agent.invoke.comfy` - Main invocation channel from router
+- `comfy.request.image` - Direct image generation requests
+- `comfy.request.video` - Direct video generation requests
+
+### Message Format
+
+**Request:**
+```json
+{
+  "type": "text-to-image",
+  "workflow": {
+    "1": {"class_type": "CLIPTextEncode", ...},
+    "2": {"class_type": "CheckpointLoaderSimple", ...}
+  }
+}
+```
+
+**Response:**
+```json
+{
+  "job_id": "job_abc123..."
+}
+```
+
+## Configuration
+
+Environment variables:
+
+- `COMFYUI_HTTP` - ComfyUI HTTP endpoint (default: `http://127.0.0.1:8188`)
+- `COMFYUI_WS` - ComfyUI WebSocket endpoint (default: `ws://127.0.0.1:8188/ws`)
+- `NATS_URL` - NATS server URL (default: `nats://144.76.224.179:4222`)
+- `STORAGE_PATH` - Path for result storage (default: `/data/comfy-results`)
+- `PUBLIC_BASE_URL` - Public URL for accessing results (default: `http://212.8.58.133:8880/files`)
+- `MAX_CONCURRENCY` - Max concurrent generations (default: `1`)
+
+## Development
+
+### Local Setup
+
+```bash
+cd services/comfy-agent
+python -m venv venv
+source venv/bin/activate  # or `venv\Scripts\activate` on Windows
+pip install -r requirements.txt
+
+# Run locally
+uvicorn app.main:app --reload --port 8880
+```
+
+### Docker Build
+
+```bash
+docker build -t comfy-agent:latest .
+```
+
+### Testing
+
+```bash
+# Test image generation
+curl -X POST http://localhost:8880/generate/image \
+  -H "Content-Type: application/json" \
+  -d '{"prompt":"a futuristic city, cyberpunk style"}'
+
+# Check status
+curl http://localhost:8880/status/job_abc123...
+
+# Health check
+curl http://localhost:8880/healthz
+```
+
+## TODO / Roadmap
+
+1. **Workflow Templates**: Replace placeholder workflows with actual ComfyUI workflows
+   - SDXL text-to-image workflow
+   - LTX-2 text-to-video workflow
+   - Image-to-video workflow
+
+2. **Result Extraction**: Implement proper file extraction from ComfyUI history
+   - Download images/videos via `/view` endpoint
+   - Support multiple output formats (PNG, JPG, GIF, MP4)
+   - Handle batch outputs
+
+3. **Advanced Features**:
+   - Workflow library management
+   - Custom model loading
+   - LoRA/ControlNet support
+   - Batch processing
+   - Queue prioritization
+
+4. **Monitoring**:
+   - Prometheus metrics
+   - Grafana dashboards
+   - Alert on failures
+   - GPU usage tracking
+
+5. **Storage**:
+   - S3/MinIO integration for scalable storage
+   - Result expiration/cleanup
+   - Thumbnail generation
+
+## Integration with Agent Registry
+
+Add to `config/agent_registry.yml`:
+
+```yaml
+comfy:
+  id: comfy
+  name: Comfy
+  role: Image & Video Generation Specialist
+  scope: node_local
+  node_id: node-3-threadripper-rtx3090
+  capabilities:
+    - text-to-image
+    - text-to-video
+    - image-to-video
+    - workflow-execution
+  api_endpoint: http://212.8.58.133:8880
+  nats_subject: agent.invoke.comfy
+```
+
+## License
+
+Part of the DAARION MicroDAO project.
+
+## Maintainers
+
+- DAARION Team
+- Last Updated: 2026-02-10
diff --git a/services/comfy-agent/app/__init__.py b/services/comfy-agent/app/__init__.py
new file mode 100644
index 00000000..f83cabd0
--- /dev/null
+++ b/services/comfy-agent/app/__init__.py
@@ -0,0 +1 @@
+# services/comfy-agent/app/__init__.py
diff --git a/services/comfy-agent/app/api.py b/services/comfy-agent/app/api.py
new file mode 100644
index 00000000..232890ae
--- /dev/null
+++ b/services/comfy-agent/app/api.py
@@ -0,0 +1,51 @@
+# services/comfy-agent/app/api.py
+from fastapi import APIRouter, HTTPException
+from .models import GenerateImageRequest, GenerateVideoRequest, JobStatus
+from .jobs import JOB_STORE
+from .worker import enqueue
+
+router = APIRouter()
+
+def _build_workflow_t2i(req: GenerateImageRequest) -> dict:
+    # MVP: placeholder graph; you will replace with your canonical Comfy workflow JSON.
+    # Keep it deterministic and param-driven.
+    return {
+        "1": {"class_type": "CLIPTextEncode", "inputs": {"text": req.prompt, "clip": ["2", 0]}},
+        "2": {"class_type": "CheckpointLoaderSimple", "inputs": {"ckpt_name": "sdxl.safetensors"}},
+        # TODO: Add complete workflow JSON for text-to-image
+    }
+
+def _build_workflow_t2v(req: GenerateVideoRequest) -> dict:
+    # MVP placeholder for LTX-2 pipeline; replace with actual LTX-2 workflow.
+    return {
+        "1": {"class_type": "CLIPTextEncode", "inputs": {"text": req.prompt, "clip": ["2", 0]}},
+        # TODO: Add complete workflow JSON for text-to-video with LTX-2
+    }
+
+@router.post("/generate/image", response_model=JobStatus)
+async def generate_image(req: GenerateImageRequest):
+    job = JOB_STORE.create("text-to-image")
+    graph = _build_workflow_t2i(req)
+    enqueue(job.job_id, "text-to-image", graph)
+    return JOB_STORE.get(job.job_id)
+
+@router.post("/generate/video", response_model=JobStatus)
+async def generate_video(req: GenerateVideoRequest):
+    job = JOB_STORE.create("text-to-video")
+    graph = _build_workflow_t2v(req)
+    enqueue(job.job_id, "text-to-video", graph)
+    return JOB_STORE.get(job.job_id)
+
+@router.get("/status/{job_id}", response_model=JobStatus)
+async def status(job_id: str):
+    job = JOB_STORE.get(job_id)
+    if not job:
+        raise HTTPException(status_code=404, detail="job_not_found")
+    return job
+
+@router.get("/result/{job_id}", response_model=JobStatus)
+async def result(job_id: str):
+    job = JOB_STORE.get(job_id)
+    if not job:
+        raise HTTPException(status_code=404, detail="job_not_found")
+    return job
diff --git a/services/comfy-agent/app/comfyui_client.py b/services/comfy-agent/app/comfyui_client.py
new file mode 100644
index 00000000..e9f7adb5
--- /dev/null
+++ b/services/comfy-agent/app/comfyui_client.py
@@ -0,0 +1,54 @@
+# services/comfy-agent/app/comfyui_client.py
+import asyncio
+import httpx
+import json
+import websockets
+from typing import Any, Dict, Optional, Callable
+from .config import settings
+
+ProgressCb = Callable[[float, str], None]
+
+class ComfyUIClient:
+    def __init__(self) -> None:
+        self.http = httpx.AsyncClient(base_url=settings.COMFYUI_HTTP, timeout=60)
+
+    async def queue_prompt(self, prompt_graph: Dict[str, Any], client_id: str) -> str:
+        # ComfyUI expects: {"prompt": {...}, "client_id": "..."}
+        r = await self.http.post("/prompt", json={"prompt": prompt_graph, "client_id": client_id})
+        r.raise_for_status()
+        data = r.json()
+        # typically returns {"prompt_id": "...", "number": ...}
+        return data["prompt_id"]
+
+    async def wait_progress(self, client_id: str, prompt_id: str, on_progress: Optional[ProgressCb] = None) -> None:
+        # WS emits progress/executing/status; keep generic handling
+        ws_url = f"{settings.COMFYUI_WS}?clientId={client_id}"
+        async with websockets.connect(ws_url, max_size=50_000_000) as ws:
+            while True:
+                msg = await ws.recv()
+                evt = json.loads(msg)
+
+                # Best-effort progress mapping
+                if evt.get("type") == "progress":
+                    data = evt.get("data", {})
+                    max_v = float(data.get("max", 1.0))
+                    val = float(data.get("value", 0.0))
+                    p = 0.0 if max_v <= 0 else min(1.0, val / max_v)
+                    if on_progress:
+                        on_progress(p, "progress")
+
+                # completion signal varies; "executing" with node=None часто означає done
+                if evt.get("type") == "executing":
+                    data = evt.get("data", {})
+                    if data.get("prompt_id") == prompt_id and data.get("node") is None:
+                        if on_progress:
+                            on_progress(1.0, "done")
+                        return
+
+    async def get_history(self, prompt_id: str) -> Dict[str, Any]:
+        r = await self.http.get(f"/history/{prompt_id}")
+        r.raise_for_status()
+        return r.json()
+
+    async def close(self) -> None:
+        await self.http.aclose()
diff --git a/services/comfy-agent/app/config.py b/services/comfy-agent/app/config.py
new file mode 100644
index 00000000..d4b5a01b
--- /dev/null
+++ b/services/comfy-agent/app/config.py
@@ -0,0 +1,22 @@
+# services/comfy-agent/app/config.py
+from pydantic_settings import BaseSettings
+
+class Settings(BaseSettings):
+    SERVICE_NAME: str = "comfy-agent"
+    API_HOST: str = "0.0.0.0"
+    API_PORT: int = 8880
+
+    COMFYUI_HTTP: str = "http://127.0.0.1:8188"
+    COMFYUI_WS: str = "ws://127.0.0.1:8188/ws"
+
+    NATS_URL: str = "nats://144.76.224.179:4222"  # NODE1 production IP
+    NATS_SUBJECT_INVOKE: str = "agent.invoke.comfy"
+    NATS_SUBJECT_IMAGE: str = "comfy.request.image"
+    NATS_SUBJECT_VIDEO: str = "comfy.request.video"
+
+    STORAGE_PATH: str = "/data/comfy-results"
+    PUBLIC_BASE_URL: str = "http://212.8.58.133:8880/files"  # NODE3 IP
+
+    MAX_CONCURRENCY: int = 1  # для LTX-2 стартово краще 1
+
+settings = Settings()
diff --git a/services/comfy-agent/app/jobs.py b/services/comfy-agent/app/jobs.py
new file mode 100644
index 00000000..272f7c46
--- /dev/null
+++ b/services/comfy-agent/app/jobs.py
@@ -0,0 +1,25 @@
+# services/comfy-agent/app/jobs.py
+import uuid
+from typing import Dict, Optional
+from .models import JobStatus, GenType
+
+class JobStore:
+    def __init__(self) -> None:
+        self._jobs: Dict[str, JobStatus] = {}
+
+    def create(self, gen_type: GenType) -> JobStatus:
+        job_id = f"job_{uuid.uuid4().hex}"
+        js = JobStatus(job_id=job_id, type=gen_type, status="queued", progress=0.0)
+        self._jobs[job_id] = js
+        return js
+
+    def get(self, job_id: str) -> Optional[JobStatus]:
+        return self._jobs.get(job_id)
+
+    def update(self, job_id: str, **patch) -> JobStatus:
+        js = self._jobs[job_id]
+        updated = js.model_copy(update=patch)
+        self._jobs[job_id] = updated
+        return updated
+
+JOB_STORE = JobStore()
diff --git a/services/comfy-agent/app/main.py b/services/comfy-agent/app/main.py
new file mode 100644
index 00000000..f7d4f504
--- /dev/null
+++ b/services/comfy-agent/app/main.py
@@ -0,0 +1,26 @@
+# services/comfy-agent/app/main.py
+import asyncio
+from fastapi import FastAPI
+from fastapi.staticfiles import StaticFiles
+from .config import settings
+from .api import router
+from .worker import worker_loop
+from .nats_client import start_nats
+from .storage import ensure_storage
+
+app = FastAPI(title="Comfy Agent Service", version="0.1.0")
+app.include_router(router)
+
+@app.on_event("startup")
+async def startup():
+    ensure_storage()
+
+    # Static files for result URLs: /files/{job_id}/...
+    app.mount("/files", StaticFiles(directory=settings.STORAGE_PATH), name="files")
+
+    asyncio.create_task(worker_loop())
+    await start_nats()
+
+@app.get("/healthz")
+async def healthz():
+    return {"ok": True, "service": settings.SERVICE_NAME}
diff --git a/services/comfy-agent/app/models.py b/services/comfy-agent/app/models.py
new file mode 100644
index 00000000..4e1001ef
--- /dev/null
+++ b/services/comfy-agent/app/models.py
@@ -0,0 +1,34 @@
+# services/comfy-agent/app/models.py
+from pydantic import BaseModel, Field
+from typing import Any, Dict, Optional, Literal
+
+GenType = Literal["text-to-image", "text-to-video", "image-to-video"]
+
+class GenerateImageRequest(BaseModel):
+    prompt: str = Field(min_length=1)
+    negative_prompt: Optional[str] = None
+    width: int = 1024
+    height: int = 1024
+    steps: int = 28
+    seed: Optional[int] = None
+    workflow: Optional[str] = None
+    workflow_params: Dict[str, Any] = Field(default_factory=dict)
+
+class GenerateVideoRequest(BaseModel):
+    prompt: str = Field(min_length=1)
+    seconds: int = 4
+    fps: int = 24
+    steps: int = 30
+    seed: Optional[int] = None
+    workflow: Optional[str] = None
+    workflow_params: Dict[str, Any] = Field(default_factory=dict)
+
+class JobStatus(BaseModel):
+    job_id: str
+    type: GenType
+    status: Literal["queued", "running", "succeeded", "failed"]
+    progress: float = 0.0
+    message: Optional[str] = None
+    result_url: Optional[str] = None
+    error: Optional[str] = None
+    comfy_prompt_id: Optional[str] = None
diff --git a/services/comfy-agent/app/nats_client.py b/services/comfy-agent/app/nats_client.py
new file mode 100644
index 00000000..bdf3afe6
--- /dev/null
+++ b/services/comfy-agent/app/nats_client.py
@@ -0,0 +1,36 @@
+# services/comfy-agent/app/nats_client.py
+import json
+import asyncio
+from nats.aio.client import Client as NATS
+from .config import settings
+from .jobs import JOB_STORE
+from .worker import enqueue
+
+async def start_nats() -> NATS:
+    nc = NATS()
+    await nc.connect(servers=[settings.NATS_URL])
+
+    async def handle(msg):
+        subj = msg.subject
+        reply = msg.reply
+        payload = json.loads(msg.data.decode("utf-8"))
+
+        # payload contract (MVP):
+        # { "type": "text-to-image|text-to-video", "workflow": {...} }
+        gen_type = payload.get("type", "text-to-image")
+        workflow = payload.get("workflow")
+        if not workflow:
+            if reply:
+                await nc.publish(reply, json.dumps({"error": "missing_workflow"}).encode())
+            return
+
+        job = JOB_STORE.create(gen_type)
+        enqueue(job.job_id, gen_type, workflow)
+
+        if reply:
+            await nc.publish(reply, json.dumps({"job_id": job.job_id}).encode())
+
+    await nc.subscribe(settings.NATS_SUBJECT_INVOKE, cb=handle)
+    await nc.subscribe(settings.NATS_SUBJECT_IMAGE, cb=handle)
+    await nc.subscribe(settings.NATS_SUBJECT_VIDEO, cb=handle)
+    return nc
diff --git a/services/comfy-agent/app/storage.py b/services/comfy-agent/app/storage.py
new file mode 100644
index 00000000..960fe6fa
--- /dev/null
+++ b/services/comfy-agent/app/storage.py
@@ -0,0 +1,16 @@
+# services/comfy-agent/app/storage.py
+import os
+from pathlib import Path
+from .config import settings
+
+def ensure_storage() -> None:
+    Path(settings.STORAGE_PATH).mkdir(parents=True, exist_ok=True)
+
+def make_job_dir(job_id: str) -> str:
+    ensure_storage()
+    d = os.path.join(settings.STORAGE_PATH, job_id)
+    Path(d).mkdir(parents=True, exist_ok=True)
+    return d
+
+def public_url(job_id: str, filename: str) -> str:
+    return f"{settings.PUBLIC_BASE_URL}/{job_id}/{filename}"
diff --git a/services/comfy-agent/app/worker.py b/services/comfy-agent/app/worker.py
new file mode 100644
index 00000000..0d9c47c3
--- /dev/null
+++ b/services/comfy-agent/app/worker.py
@@ -0,0 +1,60 @@
+# services/comfy-agent/app/worker.py
+import asyncio
+import uuid
+import os
+import json
+from typing import Any, Dict, Optional, Tuple
+from .jobs import JOB_STORE
+from .storage import make_job_dir, public_url
+from .comfyui_client import ComfyUIClient
+from .config import settings
+
+_queue: "asyncio.Queue[Tuple[str, str, Dict[str, Any]]]" = asyncio.Queue()
+
+def enqueue(job_id: str, gen_type: str, prompt_graph: Dict[str, Any]) -> None:
+    _queue.put_nowait((job_id, gen_type, prompt_graph))
+
+async def _extract_first_output(history: Dict[str, Any], job_dir: str) -> Optional[str]:
+    # ComfyUI history structure can vary; implement a conservative extraction:
+    # try to find any "images" or "gifs"/"videos" outputs and download via /view
+    # For MVP: prefer /view?filename=...&type=output&subfolder=...
+    # Here we return a "manifest.json" to unblock integration even if file fetching needs refinement.
+    manifest_path = os.path.join(job_dir, "manifest.json")
+    with open(manifest_path, "w", encoding="utf-8") as f:
+        json.dump(history, f, ensure_ascii=False, indent=2)
+    return "manifest.json"
+
+async def worker_loop() -> None:
+    client = ComfyUIClient()
+    sem = asyncio.Semaphore(settings.MAX_CONCURRENCY)
+
+    async def run_one(job_id: str, gen_type: str, prompt_graph: Dict[str, Any]) -> None:
+        async with sem:
+            JOB_STORE.update(job_id, status="running", progress=0.01)
+
+            client_id = f"comfy-agent-{uuid.uuid4().hex}"
+            def on_p(p: float, msg: str) -> None:
+                JOB_STORE.update(job_id, progress=float(p), message=msg)
+
+            try:
+                prompt_id = await client.queue_prompt(prompt_graph, client_id=client_id)
+                JOB_STORE.update(job_id, comfy_prompt_id=prompt_id)
+
+                await client.wait_progress(client_id=client_id, prompt_id=prompt_id, on_progress=on_p)
+
+                hist = await client.get_history(prompt_id)
+                job_dir = make_job_dir(job_id)
+                fname = await _extract_first_output(hist, job_dir)
+                if not fname:
+                    JOB_STORE.update(job_id, status="failed", error="No outputs found in ComfyUI history")
+                    return
+
+                url = public_url(job_id, fname)
+                JOB_STORE.update(job_id, status="succeeded", progress=1.0, result_url=url)
+
+            except Exception as e:
+                JOB_STORE.update(job_id, status="failed", error=str(e))
+
+    while True:
+        job_id, gen_type, prompt_graph = await _queue.get()
+        asyncio.create_task(run_one(job_id, gen_type, prompt_graph))
diff --git a/services/comfy-agent/requirements.txt b/services/comfy-agent/requirements.txt
new file mode 100644
index 00000000..caea14ce
--- /dev/null
+++ b/services/comfy-agent/requirements.txt
@@ -0,0 +1,9 @@
+fastapi==0.115.0
+uvicorn[standard]==0.30.6
+pydantic==2.8.2
+pydantic-settings==2.4.0
+httpx==0.27.2
+websockets==12.0
+nats-py==2.7.2
+python-multipart==0.0.9
+orjson==3.10.7