My_AI

My_AI

Your private AI assistant, running 100% on your computer

Chat, agents, documents, images and voice, powered by Ollama.
In 12 languages. Your conversations, files and memory stay on your machine.

English · Français

🚀 Quick start · ⚡ Features · 🤖 Agents · 📡 Mobile access · 🧩 VS Code extension · 📖 Documentation

✨ Why My_AI?

• 🔒 Private by design. The model runs on your PC through Ollama. Conversations, documents and memory are stored locally (ChromaDB + SQLite). Once the models are downloaded, everything except web search and mobile access works offline.
• 🧩 Everything in one window. Chat with reasoning and cited web search, 9 specialized agents and a visual workflow editor, Word, PDF, Excel and PowerPoint generation, local image generation, voice input and output, scheduled tasks.
• 📱 Wherever you work. Talk to the AI running on your PC from your phone, through an end-to-end encrypted relay, or let it read and edit your project from VS Code.
• 🌍 12 languages. The interface, the answers, the voice and the generated documents follow your language: English, French, Spanish, German, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean and Arabic (right to left).
• 🪶 Made for ordinary hardware. On first launch, the setup wizard detects your RAM and GPU, then downloads the right model (2 to 9 billion parameters). My_AI checks whether your integrated GPU runs it faster and keeps the quickest option.

🖥️ Screenshots

Agents and visual workflows	Mobile access with the Relay

🚀 Quick start

Requirements: Python 3.10+ and Ollama.

git clone https://github.com/gonicolas12/My_AI
cd My_AI
pip install -r requirements.txt

Then launch the app and choose option 1 (graphical interface):

.\launch.bat      # Windows
./launch.sh       # macOS / Linux

On the very first launch, My_AI opens in your system language (one of the 12, English otherwise, and you can change it in the wizard), detects your hardware (RAM, CPU cores, GPU VRAM), recommends and downloads the right model, then creates my_ai for you. The window opens at once on a loading screen that names each step (Ollama, model, placement, engine), then switches to the chat.

⚠️ Intel Mac (x86_64): Python 3.12 maximum. PyTorch no longer publishes x86_64 builds after version 2.2.2, whose wheels stop at Python 3.12. On Python 3.13+, the installation fails. launch.sh checks this at startup. Apple Silicon, Windows and Linux are not affected.

💡 Without a launcher, python launch_unified.py (window + Ollama setup) and python main.py --mode gui (window only, with Ollama already running) work on all three systems. On Linux, the window needs GTK or Qt (pip install "pywebview[qt]"); without them, the interface opens in your browser.

# Model (pick one for your RAM): it reads text and images
ollama pull qwen3.5:2b       # Ultra light, concise answers (4 GB RAM)
ollama pull qwen3.5:4b       # Recommended, light and fast (8 GB RAM)
ollama pull qwen3.5:9b       # Heavier, more detailed answers (16 GB RAM)

# Create the custom model
.\create_custom_model.bat

# NVIDIA (needs the NVIDIA drivers)
pip install pynvml GPUtil

# AMD (needs ROCm or the AMDGPU drivers + Microsoft C++ Build Tools)
pip install pyamdgpuinfo

# Windows, any GPU (basic info: name, total VRAM)
pip install wmi

$env:HTTP_PROXY="http://proxy.company.com:8080"
$env:HTTPS_PROXY="http://proxy.company.com:8080"
$env:NO_PROXY="localhost,127.0.0.1,::1"

export HTTP_PROXY="http://proxy.company.com:8080"
export HTTPS_PROXY="http://proxy.company.com:8080"
export NO_PROXY="localhost,127.0.0.1,::1"

.\clean_project.bat

⚡ Features

💬 Chat

🤖 AI agents

• 🧩 Custom agents: describe the agent you need, and the local model writes its system prompt.
• 🔄 Visual workflow canvas (n8n style): connectable nodes, zoom/pan, grid, minimap; sequential, parallel or DAG execution.
• 🧵 Run threads: every run is a thread you can reopen.
• 🎭 Debate mode: two agents argue a topic against each other.
• 📊 Statistics and resource monitoring: CPU, RAM, GPU, VRAM, inference time and tokens/s.
• 📅 Scheduled tasks: run an agent, a workflow or a debate on a schedule, daily, weekly, at an interval or with cron (e.g. "every morning at 8, WebAgent on AI news, then summarize"). Optionally runs even when the app is closed, through the Windows Task Scheduler, with no admin rights and no stored password. Missed tasks catch up at restart, with notifications (OS toast + in-app entry + mobile message) and .md reports in outputs/scheduled/.
• 🔔 Desktop notifications when a long task finishes.

📝 Documents

• Office documents: "generate a docx about whales", "make me a summary PDF", "create a PowerPoint presentation…", "a spreadsheet comparing…". The local model writes, and python-docx, reportlab, python-pptx and openpyxl lay it out: headings, nested lists, styled tables, Word table of contents, automatic slide splitting, Excel sheets with filters.
• In the document's language: the "Generated on" line, table of contents, page numbers and sheet names follow the language of the text. The PDF picks the right font (Chinese, Japanese, Korean, Cyrillic, Arabic).
• Editing attachments (docx, xlsx, pptx, md, pdf…): replace a text, rewrite a section, add a row or a slide. The original file is never touched: the modified copy is written to outputs/documents/.
• Native preview in the side panel, opened automatically: the Word, PowerPoint and Excel viewers of Windows Explorer, and Edge's PDF reader. HTML rendering as a fallback elsewhere, and on the mobile Relay.
• Reading files: PDF (with OCR for scanned pages), DOCX, PowerPoint, Excel, CSV and code, with smart chunking for very long documents.

🎨 Local image generation (text → image)

• Multimodal both ways: the AI sees (Ollama vision) and draws. Ask "generate an image of…", "draw me…", "create a logo…".
• Zero-config install: on the first request without a backend, My_AI downloads and starts portable ComfyUI by itself (Windows: NVIDIA, AMD or Intel Arc build depending on the GPU, otherwise on the CPU; Python + torch embedded) plus a default model, just like the first-launch wizard does for Ollama.
• 100% local through a Stable Diffusion HTTP backend: ComfyUI (auto-installed), AUTOMATIC1111 / Forge, or diffusers (optional, any GPU + CPU), set in config.yaml → image_generation:.

🎙️ Voice

• Microphone button in every input (home, conversation, Agents tab): the first click starts listening, the second one stops and transcribes, and the text is inserted at the cursor.
• 100% local with faster-whisper: small model (~150 MB) loaded on demand, language detected automatically (99+ languages supported by Whisper).
• Read aloud: a button under each answer and an automatic reading mode, with a voice per language.

🧠 Memory you control

• Vector memory of up to 10,485,760 tokens (ChromaDB + SQLite), semantic search (sentence-transformers + cosine HNSW), smart chunking with overlap (256-token chunks, 32 overlap), automatic rolling summary at half of the context window (16k tokens out of 32k).
• "Remember that…" in the chat: the fact is really saved (Facts tab of the Memory panel), and useful facts then come with every answer.
• A Memory panel to see, edit and delete what the AI knows about you, with real CRUD operations, 100% local. 3 tabs: Facts (SQLite), Documents and Conversations (ChromaDB vectors), with search, category filter, pagination and provenance (which session, which document).
• Deleting a vector entry is a real deletion in ChromaDB (collection.delete). For conversations, the "delete at the source" option removes the original message from the workspace for good.

📁 "@codebase" project context (persistent RAG)

• Attach a whole folder (a codebase or a docs folder) to a workspace: the AI keeps that context available for every question, with no need to drop the files again.
• 100% local incremental indexing: reuses the vector memory (codebase collection) and the existing processors; respects .gitignore and skips node_modules, .git, .venv, dist… Only changed files are processed again when reindexing.
• Retrieval at question time (RAG), plus a deterministic shortcut for questions about the folder (path, list of files).
• Window: "Project folders" section of the sidebar. VS Code: the @ menu attaches files and folders, one @codebase context per project.

🔌 MCP and full PC access

• MCP (Model Context Protocol) integration: standardized connection to local tools and external servers (files, git, databases).
• Local tools to read, write and move files and create folders, with the full paths returned by the tools checked carefully, so the AI can organize workspaces on its own.
• A confirmation dialog before any file deletion.

🦙 Ollama integration: 100% local LLM

• Multiple models: qwen3.5:2b, qwen3.5:4b, qwen3.5:9b (text and images), mistral...; a dedicated vision model (minicpm-v, llava, llama3.2-vision...) handles images only when the chosen model can't see them.
• Automatic placement: at launch, My_AI checks where Ollama puts the model. If it stays on the processor, My_AI measures whether the integrated GPU is faster and keeps the quicker one; when the model overflows the graphics card (or the integrated GPU), its context switches to 8 bits if that is faster. The choice is kept, and measured again after an Ollama update or a hardware change.
• A single context window (32,768 tokens, 65,536 as an option) for every call: Ollama no longer reloads the model from one step to the next and resumes the conversation it has already read.
• Customizable model: temperature, context and system prompt through the Modelfile.
• Smart fallback: without Ollama, the AI automatically switches to a patterns/rules mode.

🌐 And also

• 🌐 REST API: built-in FastAPI server for external integrations.
• 🎓 RLHF, training pipeline and compression: learning from feedback, fine-tuning with monitoring, detailed compression ratios (advanced features).
• 💻 Cross-platform: Windows · macOS · Linux.

📡 My_AI Relay — mobile access

Talk to your AI from your phone (iOS/Android), wherever you are, as long as the app is running on your PC.

How it works

1. Click Relay at the bottom of the sidebar (or Ctrl+R)
2. My_AI Relay starts a WebSocket server and opens secure tunnels in parallel (cloudflared and localhost.run)
3. Scan the QR code or copy the URL to your phone, then enter the access code shown under the QR code
4. The mobile interface has two tabs at the top, 💬 Chat and 🤖 Agents (like the desktop app):
   • Chat: chat in real time; messages also appear on the PC. The send button turns into a Stop button during generation, and [title](url) links show up as clickable blue links.
   • Agents: the whole Agents page of the PC, made for touch: agent grid, creating/editing/deleting custom agents, n8n-style visual workflow (drag nodes, link ports), Debate mode and streamed runs.
5. Attach images, PDF, DOCX, PowerPoint, Excel or code files with the + button (in the Chat and on the Agents page): they go through the same processors as on the PC (vision model for images, vector context for documents)
6. Ask for a document ("generate a pdf about…"): its preview opens automatically in the modal, with a 💾 button to download the file

Features

Configuration (config.yaml)

relay:
  auto_start: false      # Start automatically with the GUI
  port: 8765             # Relay server port
  response_timeout: 500  # Maximum AI response time (seconds)
  tunnel: true           # Enable the tunnels (tunnel_providers)
  tunnel_providers: ["cloudflared", "localhost.run"]  # "serveo" as an option, see docs/USAGE.md
  host: "0.0.0.0"        # Listening address

Without a tunnel (tunnel: false, or no reachable provider), the Relay is only available on your local network.

🧩 VS Code extension

My_AI Relay is also available as an extension on the VS Code Marketplace. It does more than relay the chat: it offers an agentic mode where the local LLM (on the host PC) can read, edit and create files, run shell commands and search your VS Code workspace, with every action visible and approvable in the chat.

How it works

1. On the host PC: start the Relay (Relay at the bottom of the sidebar, or Ctrl+R) → as soon as a tunnel answers, click Copy in the panel: the connection string is the QR code link
2. In VS Code: install the My_AI Relay extension from the Marketplace → open its view in the sidebar → Paste the connection string → enter the access code shown under the QR code
3. Once connected, the extension identifies itself as a vscode client to the Relay and turns on the agentic mode. Ask for example "scaffold a Vite + React project with a router and a Header component": you will see each write_file / run_command appear as an expandable card in the chat, with an approval request for destructive operations.

One Code conversation, three places. The VS Code conversation also shows on the Code page of the desktop app (Code button at the top of the sidebar) and in the Code view of the phone (</> button left of Chat): messages typed there go to the extension, which acts in its workspace, and the reply streams everywhere. On the desktop page, messages can carry the chat's attachments (documents, code, images described by the vision model) and be dictated with the local voice input. A write or a shell command asked from the PC or the phone waits for an approval that can be given from any of the three. The regular chat (phone + desktop) stays separate: nothing of it reaches VS Code, and the other way round. The regular chat keeps the classic pipeline with the full local MCP tools: the agentic mode is exclusive to the Code conversation.

Features

Installation

# From VS Code: Extensions (Ctrl+Shift+X) → search "My_AI Relay" → Install
# OR from the command line:
code --install-extension gonicolas12.my-ai

See vscode_extension/README.md (English) or vscode_extension/README.fr.md (French) for the full documentation, the complete list of tools and the details of the approval flow.

📖 Documentation

The detailed guides are written in French for now.

🏗️ Architecture

My_AI/
├── core/              # AI engine, chat and agent orchestration, memory, scheduler, REST API
├── frontend/          # Web interface (React + TypeScript, Vite), built into frontend/dist
├── interfaces/        # Native window (pywebview) and its local server, CLI
├── models/            # Agents, Ollama manager, web search, image generation
├── processors/        # Readers for PDF (OCR), DOCX, PowerPoint, Excel and code
├── generators/        # Word/PDF/PowerPoint/Excel generation and editing
├── memory/            # Vector memory (ChromaDB)
├── relay/             # My_AI Relay: mobile access server and PWA
├── vscode_extension/  # VS Code extension (Relay client, agentic mode)
├── utils/ · tools/    # Utilities, model placement, cloudflared, demo server
├── main.py · launch_unified.py · launch.bat · launch.sh
└── config.yaml · Modelfile

The file-by-file tree and the processing flows are in docs/ARCHITECTURE.md.

🗺️ Roadmap

• 🔊 Voice: voice output on mobile/Relay and a hands-free conversation mode (dictation → answer → read aloud, in a loop)
• 🧩 VS Code extension: integration with VS Code diagnostics (Problems panel), before/after diffs for edits, dedicated terminals for run_command
• 🧠 Better reasoning engine (Thinking mode)
• 🔗 Third-party API integrations

🤝 Contributing

Issues and pull requests are welcome. For a large change, please open an issue first so we can discuss it. A bug report with steps to reproduce helps a lot.

If My_AI is useful to you, a ⭐ on GitHub helps other people find it.

📄 License

MIT

Built with ❤️ to stay local, private and powerful.