Neuralforge

🧠 NeuralForge

Self-hosted AI command center. 11 services. 69 APIs. Zero cloud.

LLM agents, SMM autopilot for 7 platforms, image/video/3D/music generation,RAG, LoRA fine-tuning, voice cloning, Telegram bot with vision — all from localhost:9000

---

> Why another AI dashboard? Because no other open-source project gives you LLM orchestration, automated SMM for 7 platforms, image/video/3D generation, RAG, fine-tuning, Telegram bot with 14 personas, voice cloning, and MCP integration for Claude — all in a single self-hosted panel with zero cloud dependencies.

---

Highlights

• 🤖 11 AI Services managed from one UI — Ollama, ComfyUI, Whisper, Qdrant, SearXNG, and more
• 📱 SMM AI Department — discover trends → generate posts → create images → auto-publish to Telegram, Twitter, Facebook, Instagram, Threads, LinkedIn, Discord simultaneously
• 🧠 Multi-Agent System — 13 roles, 3 modes (Solo/Team/Orchestrator), 9 tools including web search, code execution, RAG
• 🎨 Full Generation Pipeline — Image (FLUX) → Video (Wan2.2) → 3D (Hunyuan3D) with smart VRAM management
• 📊 69 API Endpoints — everything is programmable, extensible, and automatable
• 🔒 100% Local — your data never leaves your machine. No API keys required for core features

---

Table of Contents

[What is this?](#what-is-this) · [AI Model Stack](#ai-model-stack) · [Features](#features-at-a-glance) · [Dashboard](#dashboard) · [Agents](#ai-agents) · [RAG](#rag-retrieval-augmented-generation) · [LoRA](#lora-fine-tuning) · [Pipeline](#generation-pipeline) · [Telegram Bot](#telegram-ai-bot) · [SMM](#smm-ai-department) · [MCP Server](#mcp-server--claude-code-integration) · [Quick Start](#quick-start) · [Requirements](#requirements) · [FAQ](#faq)

---

What is this?

A self-hosted web panel (localhost:9000) that unifies your entire local AI infrastructure into one powerful dashboard. No subscriptions, no cloud APIs, no data leaving your machine.

┌──────────────────────────────── NeuralForge ──────────────────────────────────┐
│                                                                                │
│  Dashboard     Agents       RAG        Telegram     LoRA        SMM           │
│  ┌────────┐   ┌────────┐  ┌────────┐  ┌────────┐  ┌────────┐  ┌────────┐    │
│  │GPU/VRAM│   │13 Roles│  │Qdrant +│  │14 Meme │  │Unsloth │  │7 Socials│   │
│  │Services│   │Solo    │  │ONNX GPU│  │Personas│  │LoRA    │  │Trend AI │   │
│  │Metrics │   │Team    │  │1800/sec│  │Voice   │  │16 base │  │Post Gen │   │
│  │Alerts  │   │Orchestr│  │Multi-DB│  │Cloning │  │models  │  │Analytics│   │
│  └────────┘   └────────┘  └────────┘  └────────┘  └────────┘  └────────┘    │
│                                                                                │
│  Pipeline: Image ──→ Video ──→ 3D   │   MCP Server: 24 tools for Claude      │
│  (ComfyUI)  (Wan2GP)  (Hunyuan3D)   │   + Music, TTS, STT, Search...        │
└────────────────────────────────────────────────────────────────────────────────┘

AI Model Stack

Every model runs locally via Ollama — no API keys, no cloud, no subscriptions.

LLMs (Text Generation & Reasoning)

| Model | Size | VRAM | Used for | |-------|------|------|----------| | Qwen 3.5 | 35B (A3B MoE) | ~20 GB | Primary workhorse — agents, SMM posts, trend analysis | | Nemotron 3 Nano | 30B | ~18 GB | RAG answers, balanced quality/speed | | Mistral Small | 24B | ~14 GB | Summarization, translation, email | | Qwen 3.5 | 9B | ~6 GB | Telegram bot — fast persona responses | | Gemma 3 | 27B | ~16 GB | Alternative general-purpose | | DeepSeek R1 | 14B | ~9 GB | Math, reasoning, code | | + 9 more | 1B–35B | 1–20 GB | User-selectable per task |

Vision (Image Understanding)

| Model | Size | VRAM | Used for | |-------|------|------|----------| | MiniCPM-V | 8B | ~5 GB | Telegram bot photo analysis — describes images, answers questions about photos sent to your account | | Qwen2.5-VL | 27B | ~16 GB | Agent image analysis tool — detailed visual Q&A |

Embeddings (RAG Search)

| Model | Size | Speed | Used for | |-------|------|-------|----------| | bge-m3 (ONNX) | 560M | 1,800 docs/sec | GPU-accelerated document indexing | | bge-m3 (Ollama) | 560M | 10 docs/sec | Fallback CPU embedding |

Audio (Speech & Music)

| Model | Size | VRAM | Used for | |-------|------|------|----------| | Whisper (faster-whisper) | base/large | 2-10 GB | Speech-to-text, 99 languages, diarization | | Qwen3-TTS | ~4 GB | ~4 GB | Text-to-speech, 3-second voice cloning | | ACE-Step 1.5 | ~4 GB | 4-6 GB | AI music generation — lyrics + style → full song |

Image / Video / 3D Generation

| Model | Engine | VRAM | Used for | |-------|--------|------|----------| | FLUX Klein | ComfyUI | 8-12 GB | Image generation (SMM posts, pipeline) | | Wan 2.2 | Wan2GP | 12-24 GB | Video generation from image + prompt | | Hunyuan3D v2 | Gradio | 13-20 GB | 3D model generation from image |

LoRA Fine-Tuning (16 base models)

| Model | Size | Training time | |-------|------|---------------| | NVIDIA Nemotron 3 Nano | 4B | ~30 min | | Llama 3.1 / 3.2 | 1B–8B | 30 min – 2h | | Qwen 2.5 | 7B / 32B | 1–4h | | Gemma 2 | 2B–27B | 30 min – 3h | | Mistral v0.3 | 7B | ~1h | | Phi 3.5 | 3.8B | ~40 min |

> Total unique AI models available: 30+ — all running locally, swappable per task, with automatic VRAM management.

Features at a Glance

| | Feature | Description | |:---:|---|---| | GPU | Smart VRAM Management | Exclusive groups auto-stop conflicting services. Never OOM again | | Dashboard | Real-time Monitoring | GPU temp, VRAM, RAM, CPU, disk — live metrics with health alerts | | Agents | Multi-Agent Orchestration | 13 roles, 3 modes (Solo/Team/Orchestrator), shared memory, RAG tools | | RAG | Vector Search at GPU Speed | ONNX embeddings at 1,800 docs/sec, Qdrant DB, multi-collection search | | Bot | 14 Telegram Personas | Each with unique personality — from Philosopher to Crypto Maniac | | Voice | Real-time Voice Cloning | Send voice → get reply in your own voice with AI-generated text | | LoRA | Fine-Tuning UI | 16 base models, dataset upload, live training output, adapter export | | Gen | Image→Video→3D Pipeline | Automated chain with smart VRAM switching between steps | | MCP | Claude Code Integration | 24 tools — let Claude manage your entire AI stack | | SMM | 7-Platform Social Media | Trend Scout → AI Post Writer → Image Gen → Auto-Publish to all platforms | | Ext | YAML Module System | Add any new service in 10 lines of YAML |

Dashboard

The main hub. Everything starts here.

Live Metrics:

• GPU VRAM usage with free memory indicator
• GPU temperature and power draw
• RAM usage with available memory
• CPU load across all threads
• Disk usage with free space alerts

Service Management:

• Start/stop any service with one click
• Exclusive GPU groups — when you start ComfyUI, Wan2GP auto-stops (and vice versa). No more VRAM crashes
• Service health indicators (running/stopped/starting)
• Quick actions: "Start basics", "Stop heavy", "Free VRAM"

Monitoring:

• Active Ollama models with per-model VRAM usage
• GPU process list (what's eating your VRAM right now)
• Qdrant RAG collections with vector counts
• Storage breakdown by service (ComfyUI outputs, Wan2GP videos, etc.)
• Health alerts: GPU overheating, low disk, service down — all visible at a glance

YAML Module System — add any service:

name: My New Service
category: generation
start_cmd: "python3 app.py --port 7777"
port: 7777
vram_estimate: "4-8 GB"
exclusive_group: heavy_gpu    # auto-stops conflicting services

Drop it in modules/ → restart panel → it appears. That's it.

AI Agents

A full multi-agent framework built into the panel.

13 Role Presets:

| Role | What it does | Default model | |------|-------------|---------------| | Researcher | Web search, source analysis, fact compilation | Qwen 3.5 35B | | Analyst | Data analysis, pattern recognition, insights | Qwen 3.5 35B | | Coder | Write, debug, refactor code in any language | Qwen 3.5 35B | | Writer | Articles, reports, creative writing | Qwen 3.5 35B | | Critic | Quality review, scoring, improvement suggestions | Qwen 3.5 35B | | Summarizer | Condense long texts into key points | Mistral Small 24B | | Translator | Multi-language translation with context | Mistral Small 24B | | Email Writer | Professional emails from brief instructions | Mistral Small 24B | | Tester | Generate test cases, find edge cases | Qwen 3.5 35B | | Trade Analyst | Market analysis, trend identification | Qwen 3.5 35B | | Tutor | Explain concepts at adjustable complexity | Qwen 3.5 35B | | Security Auditor | Code/config security review, vulnerability scan | Qwen 3.5 35B | | Image Analyst | Describe and analyze images | Qwen Vision 27B |

3 Execution Modes:

| Mode | How it works | Best for | |------|-------------|----------| | Solo | Single agent with tools | Quick tasks, Q&A | | Team | Agent chain — each passes context to next | Complex multi-step tasks | | Orchestrator | AI creates plan → delegates to agents → reviews result (retries if score

RAG (Retrieval-Augmented Generation)

Ask questions about your documents. The AI retrieves relevant passages and answers with citations.

Performance: | Method | Speed | GPU VRAM | |--------|-------|----------| | ONNX GPU (bge-m3) | 1,800 texts/sec | ~2 GB | | Ollama embeddings | 10 texts/sec | ~4 GB |

That's 180x faster indexing with ONNX.

Capabilities:

• Multi-format indexing — PDF, TXT, MD, DOCX, CSV, HTML
• Batch processing — index entire directories recursively
• Multi-collection — separate databases for different topics (e.g., "laws", "docs", "codebase")
• Smart search — auto-detects which collection to search based on query
• Context memory — remembers previous Q&A in the same chat session
• Embedding cache — repeat queries are instant

Built-in chat interface:

• Markdown rendering with syntax highlighting
• Copy button on every response
• Export conversation to Markdown file
• Collection selector and search settings
• localStorage persistence — your chat survives page reload

Example use case: > Indexed all 390 Estonian laws (52,314 vectors) — now ask legal questions in any language and get answers with article references.

LoRA Fine-Tuning

Train custom model adapters directly from the panel UI.

16 Base Models Ready to Fine-Tune:

| Model | Size | Notes | |-------|------|-------| | Llama 3.1 | 8B | Great all-rounder | | Llama 3.2 | 1B / 3B | Lightweight, fast | | Mistral v0.3 | 7B | Strong reasoning | | Qwen 2.5 | 7B / 32B | Multilingual | | Gemma 2 | 2B / 9B / 27B | Google's latest | | Phi 3.5 | 3.8B | Microsoft, compact | | + custom | any | Enter any Unsloth-compatible model ID |

Training UI Features:

• Dataset upload (JSON, JSONL, CSV) or HuggingFace dataset ID
• Auto-format detection (instruction/output, messages, or raw text)
• Configurable: LoRA rank, alpha, epochs, batch size, learning rate, max sequence length
• Live training output — see loss, progress, ETA in real-time
• Timer showing elapsed training time
• Stop button to cancel mid-training
• Trained adapters listed with size and date

Powered by Unsloth — 2x faster training, 60% less memory than standard LoRA.

Generation Pipeline

Automated Image → Video → 3D chain with smart VRAM management between steps.

| Step | Engine | VRAM | Automation | |------|--------|------|------------| | Image | ComfyUI (FLUX Klein 4B) | 8-12 GB | Fully automated API | | Video | Wan2GP (Wan 2.2 / LTX) | 12-24 GB | Gradio API + manual fallback | | 3D | Hunyuan3D | 13-20 GB | Gradio API + manual fallback |

VRAM is automatically freed between steps — only one heavy service runs at a time.

# 5 built-in examples
python3 pipeline.py --example robot     # chibi robot → animate → 3D model
python3 pipeline.py --example dragon    # crystal dragon → animate → 3D
python3 pipeline.py --example car       # cyberpunk car → animate → 3D
python3 pipeline.py --example cat       # cat astronaut → animate → 3D
python3 pipeline.py --example sword     # magic sword → 3D (skip video)

# Custom prompt
python3 pipeline.py "a golden crown with gems" --steps image,3d
python3 pipeline.py "a phoenix" --video-prompt "spreads wings and flies"

Telegram AI Bot

Not a Telegram bot — responds from your own account via Telethon User API.

14 Unique Personas:

| | Persona | Style | |:---:|---|---| | 🧘 | Philosopher | "You wrote 'hi', but what is a greeting if not a scream of loneliness into the void?" | | 🧢 | Street Philosopher | "bro, your argument is logically inconsistent, purely by Kant" | | 👾 | IT Demon | "segfault in your logic, recompile that thought" | | 👵 | Granny from 2077 | "sweetie, browsing without a firewall again? you'll catch a virus!" | | 🕵️ | Noir Detective | "The message came at 3am. Like all bad news in this city" | | 🏴‍☠️ | Nerd Pirate | "arrr, your meme is a true treasure!" | | 🐱 | Cat Overlord | "I'd help, but I need to lie down for 14 more hours" | | 🔺 | Conspiracy Nut | "Telegram was created by Masons to track memes" | | 🎭 | Budget Shakespeare | "To be online or not to be — that is the question!" | | 🧟 | Polite Zombie | "good evening, could you... share some brains?" | | 📋 | Corporate Robot | "let's sync on this in the next sprint" | | 🫎 | Capybara | "why stress when you can just... not" + random capybara photo | | 🚀 | Crypto Maniac | "RED CANDLE, I'M BANKRUPT, wait... GREEN! I'M RICH!" | | 🛠️ | Custom | Write your own character |

Voice Clone Pipeline:

🎤 Voice in → ffmpeg (OGG→WAV) → Whisper STT → LLM response
  → unload LLM → Qwen3-TTS (clone voice) → ffmpeg (WAV→OGG) → 🔊 Voice out

Vision — Photo Analysis (MiniCPM-V 8B):

📸 Photo in → MiniCPM-V (image description) → LLM (persona-styled response) → 💬 Reply

Send a photo to your account → the bot describes it through the vision model → responds in character. Toggle on/off from panel UI.

Features:

• Vision mode — understands photos via MiniCPM-V 8B (auto VRAM swap: unload LLM → load vision → analyze → unload → reload LLM)
• Auto-detects language → responds in same language
• Conversation memory (5 exchanges per user)
• Session-based logs grouped by contact
• Voice clone toggle from panel UI
• Capybara persona sends random capybara photos via capy.lol API

SMM AI Department

Fully automated social media management system — from trend discovery to publishing across 7 platforms.

Complete Workflow:

Trend Scout → Post Writer → Image Gen → Content Queue → Auto-Publish
    │              │             │              │              │
    ▼              ▼             ▼              ▼              ▼
 6 Sources     2-Pass LLM    ComfyUI FLUX   Schedule +     7 Platforms
 (Reddit,HN,  (Scrape→       + ffmpeg      Calendar       simultaneously
  GitHub,RSS,  Summary→       resize        view           with retry
  SearXNG,     Platform
  GoogTrends)  posts)

7 Connected Platforms: | Platform | Auth Method | Features | |---|---|---| | Telegram | Bot API | Text + Photo, channel posting | | Discord | Webhook | Text + File upload | | Twitter/X | OAuth 1.0a | Text + Media upload (Pay-Per-Use) | | Facebook | Page Token (permanent) | Text + Photo, Page posting | | Instagram | Graph API via FB | Photo + Caption (via imgur) | | Threads | Threads API | Text + Image | | LinkedIn | OAuth 2.0 | Text + Image (3-step upload) |

Key Features:

• Trend Scout v2 — Multi-source intelligence with nich

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: DefinitelyN0tMe
• Source: DefinitelyN0tMe/neuralforge
• License: MIT

Install and usage instructions live in the source repository linked above.