TAI is an integrated suite of tools designed for AI-driven software development, autonomous automation, and programmable intelligence. It treats software development as a process of Theory Building, providing the infrastructure to bridge high-level intent with concrete execution.
At its core, TAI combines advanced AI model orchestration with specialized virtual machines and autonomous agents to create a highly capable development environment.
TAI includes two specialized execution environments that allow AI models to run code safely and deterministically:
- Taipy: A Pythonic scripting language (Starlark-based syntax) compiled to custom bytecode.
- Taigo: A lightweight Go interpreter that supports a significant subset of Go syntax for dynamic logic execution.
taido is a minimalist autonomous execution engine. It implements a ReAct (Thought-Action-Observation) loop that allows an AI agent to:
- Reason about a goal.
- Act using tools (Shell, EvalTaigo, Sub-agents).
- Observe outputs and adjust its strategy. It is designed for "set and forget" tasks like iterative optimization, data processing, or environment setup.
A unified interface for multiple AI providers, including:
- Google Gemini (Native support for thinking/reasoning models).
- OpenAI & Compatible Providers (Deepseek, OpenRouter, Baidu, Tencent, Aliyun, etc.).
- Support for structured output, tool calling, and token counting across different architectures.
TAI provides several commands for different workflows:
tai chat: An interactive AI assistant with a persistent memory system (User Profile) and tool-calling capabilities.tai do: Triggerstaidofor autonomous task execution. Example:tai do goal="Refactor the storage layer to use PostgreSQL".tai go: AI-powered Go code analysis and transformation tool.tai any: General-purpose AI coding assistant for any language.taipatch: Applies code hunks (in TAI's custom format) generated by the AI to the local filesystem safely.
TAI uses a hierarchical, dual-layered configuration system:
- Data Configuration (
tai.cue): Uses CUE for type-safe, schema-validated settings (API keys, model preferences, file filters). - Logic Configuration (
tai.go): UsesTaigoto allow programmable configuration. You can define variables and logic in Go that are accessible to the AI during its reasoning process.
Configuration is loaded in order: /etc/tai.go -> ~/.config/tai.go -> ./tai.go.
TAI prioritizes safety when giving AI models access to your system:
- Sandboxing: On Linux,
taidosupports an opt-in filesystem sandbox using Landlock. When enabled (via-safe), the agent is restricted to write access only within the current working directory while maintaining read access for context. - Containerization:
gotaican automatically re-execute itself within a restricted namespace/container to provide further isolation.
go install github.com/reusee/tai/cmd/tai@latestStart a chat with memory enabled:
tai chat "How does the project's VM handle closures?"Execute an autonomous task:
tai do goal="Find all TODOs in the codebase and summarize them in TODO.md"The project follows the philosophy of Conceptual Integrity. Instead of relying on complex, opaque frameworks, TAI provides lean abstractions that maintain a clear "Theory of the System."
- Software Development as Theory Building: TAI is built to help users and AI build a shared mental model of the code.
- Strategic Subtraction: We prefer minimal, composable tools over monolithic features.
- Interface First: Clear semantics for tool usage and model interaction.
| Directory | Description |
|---|---|
cmd/ |
CLI tool implementations (tai, taigo, taipy, etc.) |
taivm/ |
The core stack-based Virtual Machine |
taipy/ |
Pythonic compiler and language bindings |
taigo/ |
Go interpreter and compiler |
taido/ |
Autonomous execution logic |
generators/ |
AI model provider implementations |
anytexts/ |
Context providers for various file types |
gocodes/ |
Specialized Go source analysis and simplification |