Tao AI Studio is a desktop IDE that orchestrates multiple AI models simultaneously — with persistent memory, parallel agents, document editing, and tool calling built in from the ground up. It ships as a single EXE with no complex installation — just .NET 9 is required, which comes standard with Windows 11.
Windows 10/11 · .NET 9 Runtime included · Free
From code editing and AI chat to office document generation and autonomous agents — all in a single desktop application.
Persistent episodic memory with vector search. Memories are organised into Rooms, Halls, Chains, and Tunnels — surfacing the right context every time.
Spawn independent background worker agents that run up to 40-turn agentic loops concurrently. Monitor live progress without blocking the main chat.
Edit spreadsheets with natural language. Type "Create a budget tracker" and watch the AI generate formulas, data, and formatting automatically.
Automatic self-review loop that keeps agents accountable. The agent evaluates its own work against project tracker goals and continues until satisfied.
200+ available tools organised into categories that auto-expire after 10 seconds of inactivity, keeping every LLM call lean and focused.
Switch between Ollama, OpenRouter, vLLM, LocalAI, HuggingFace, and Llama.cpp with a model prefix. Unified message format across all providers.
Full-featured AvalonEdit integration with syntax highlighting, multi-tab support, and automatic file type detection across 15+ languages.
Generate class, sequence, component, and namespace summary diagrams. Pre-generated diagrams can be injected into the system prompt for architectural awareness.
Vision model support for image analysis plus text-to-image generation. Attach images, videos, audio, and documents as conversation context.
Transform any image into a 3D model using HuggingFace Trellis or Stability AI's Stable Fast 3D. Output GLB files and view them in the built-in Three.js 3D viewer tab.
Model Context Protocol integration — connect any MCP server and its tools are automatically discovered and merged with the built-in tool library.
Enable reasoning mode on supported models with a configurable 1K–32K token budget. Works with OpenRouter thinking models, vLLM, and llama.cpp /think commands.
A structured planning interface lets the agent outline its approach before executing. Working memory gives agents a persistent scratch pad across turns.
Integrated Brave Search API gives agents real-time internet access with 2,000 free calls/month. Paired with a web-fetch tool for full page content extraction.
9 email and calendar tools — read unread mail, send emails with attachments, create appointments, search inbox, and manage contacts via Microsoft Outlook.
Open any file and the right editor loads automatically. Every document type gets its own AI-integrated experience.
AvalonEdit code editor with syntax highlighting for 15+ languages
WebView2 chat with markdown rendering and streaming responses
100-turn autonomous agent with full tool access and Goal Check
Display generated and processed images; attach to chat for vision analysis
Playback MP4, AVI, MOV, MKV, WEBM directly in a tab
WAV, MP3, M4A, OGG, FLAC playback plus AI speech synthesis output
DataGrid spreadsheet with inline AI chat for natural language editing
Document tab with integrated AI chat for content generation
Slides creation and editing with AI chat assistance
View and interact with PDF documents with integrated AI chat
Rendered Mermaid diagrams — class, sequence, component, namespace
Three.js GLB viewer with pan, zoom, and rotate — outputs from Trellis/Stable Fast 3D
Embedded browser tab with back/forward/refresh navigation controls
Render and inspect SVG drawings and vector graphics inline
Tao AI Studio automatically detects file type on open and routes to the correct tab — code files open in AvalonEdit, spreadsheets launch the Spreadsheet+AI tab, images appear in the viewer, and media files play immediately.
Open any .html or .md file as raw text in the code editor. Click the filename or the eye icon in the tab bar and the file instantly opens in its dedicated preview tab — no manual conversion needed.
Switch between local servers and cloud APIs with a simple model prefix. A single unified message format keeps your code provider-agnostic.
Local LLM server with native concurrent client support. The default provider — zero config needed for most setups.
Cloud gateway to hundreds of frontier models. Real-time cost and token tracking per message. Extended thinking supported with 1K–32K token budgets.
High-throughput inference server. Configurable endpoint in settings — ideal for quantised open-weight models at scale.
Self-hosted OpenAI-compatible server. Single-process — the agent semaphore serialises worker calls automatically.
Inference API integration supporting thousands of models on HF Hub. Requires an API key configured in Settings.
Direct llama.cpp server integration for maximum control over quantisation and GPU offloading on consumer hardware.
Inspired by the Method of Loci, the Memory Palace organises knowledge spatially. Every conversation is encoded, clustered, linked, and retrievable in milliseconds — with zero extra VRAM.
Auto-clustered topic groups. Each room has a generated label from anchor keywords.
Knowledge classification: fact · procedure · event · decision
Episode links via PreviousMemoryId — walk the chain for full conversation replay.
Cross-topic bridges at 0.65 similarity — surface related ideas from other projects.
The default CurrentModel strategy reuses the active chat model for embeddings — no second model to load, no extra GPU memory consumed. Switch to a dedicated encoder for higher precision, or to keyword TF-IDF for fully offline operation.
The root agent runs up to 100 turns autonomously and can spawn background workers — each running their own independent 40-turn agentic loop. The root never blocks while workers run in parallel.
launch_parallel_agent
Each worker writes its state to .tao/agents/{name}/. The AgentStatusDialog uses a FileSystemWatcher for instant OS-level notifications — no polling, no timers.
Each provider gets a dedicated service instance per worker — no shared event handlers, no LLM response races. A single semaphore serialises calls to single-process servers like vLLM and LocalAI.
Enable Goal Check and the agent is held accountable to your active Project Tracker goal. When it stops, it must evaluate its own work — and keeps going until it's genuinely satisfied.
No tool calls detected — the agent appears to be done with its current response.
A structured review message referencing the active goal is injected as a user message — covering quality, testing, output verification, and blockers.
completed()satisfied: true — stop. satisfied: false — continue autonomously until the next review.
The cycle repeats each time the agent stops — guaranteeing it doesn't declare victory prematurely.
The agent can autonomously create, populate, verify, and screenshot Office documents — all via tool calls, with tab locking to prevent interference.
"Create a monthly budget tracker with formulas"
Current spreadsheet state + message → JSON operations from LLM
State manager applies cell edits, formulas, and formatting
DataGrid updates instantly — iterate with follow-up messages
The OfficeTabs tool category lets agents open, command, and verify Office documents. Tab locking prevents the user from interfering while the agent works.
Agents can create polished documents and presentations from scratch. Each tab includes an embedded AI chat — ask follow-up questions, request edits, or have the agent iterate on content.
Drop any 2D image and Tao AI Studio converts it to a production-ready 3D GLB model via HuggingFace Gradio Spaces — no local GPU required. The result opens instantly in the built-in Three.js viewer.
JeffreyXiang/TRELLIS — highest quality multi-view 3D reconstruction. Best for objects and products.
stabilityai/stable-fast-3d — ~30 second conversion via Stability AI's HuggingFace Space. Fast previews.
Tool categories auto-disable after 10 seconds of inactivity. Every LLM turn gets only the tools it needs — keeping context windows lean and costs down. Connect MCP servers to add more.
Core tools always available: completed(), enable_tool_category, working memory, plan mode, and agent management.
Read, write, glob, zip, hash. Directory ops, file moves, and properties. Auto-expires 10s after last use.
Brave Search API, web fetch with content extraction, URL accessibility checks, and browser bookmarks.
Launch apps, screen capture, keystrokes, mouse control, clipboard, registry, network, system monitoring.
Open and control spreadsheet, document, and presentation tabs. Screenshot verification included.
Goals, tasks, subtasks — plus pt_git_* tools for commit, push, diff, log, and branch from within the tracker.
Search by topic, date, or recency. Retrieve statistics and store new episodic memories mid-conversation.
Launch, poll, retrieve output from, and kill parallel worker agents. Root-agent only.
Read unread mail, send with attachments, create calendar appointments, search inbox, manage contacts.
Tools from any connected Model Context Protocol server are auto-discovered and added to the active set.
Convert 2D images to GLB models via Trellis or Stable Fast 3D. Open output in the built-in Three.js viewer.
Inject images into the next turn with see_image, capture screen regions, and screenshot any open tab.
Built on .NET 9 WPF with battle-tested libraries for every capability.
Full-featured code editor component with syntax highlighting, code folding, and line numbers.
Microsoft's Chromium-based WebView for rendering AI chat with full markdown and HTML support.
Pure .NET spreadsheet library for creating and modifying spreadsheets — no Office installation required.
Embedded document database for persistent memory storage — zero config, serverless.
Official .NET client for Ollama with streaming, tool calling, and model management.
.NET speech synthesis for AI response audio playback — text-to-speech built into .NET 9.
Build in Debug for development or Release for distribution. Open in Visual Studio for the full WPF design experience.