wtf is taicode?

TAICode is a learning project that builds an AI coding agent modeled after tools like Claude Code. It is built on LangGraph, a framework for building AI agents as state machines, combined with a reasoning approach called ReAct and task graphs that let work run in parallel. The project supports memory across sessions, the Model Context Protocol for connecting external tools, reusable skills, coordination between multiple agents, and both a terminal interface and a web interface. When a user sends a request, a routing layer decides whether it is simple enough for a single worker to handle directly or complex enough to go through a planner. The planner first inspects the project using tools like listing files, reading files, and searching text, then breaks the task into a dependency graph so that up to eight workers can execute pieces of it at the same time once their dependencies are satisfied. A separate context compiler continuously scores and trims the information being sent to the model so that long tasks do not overwhelm it with stale or low-value context. Memory is organized into four layers: short-term memory, semantic memory, a user profile, and stored rules, all converted into vector representations locally using a small embedding model so nothing needs to be sent to an external service just to search memory. Skills work similarly, with markdown files scanned from a local folder, converted to embeddings, and automatically matched to relevant tasks using similarity search, so a long list of instructions does not need to be pasted into every prompt by hand. Built-in tools cover common file and shell operations such as reading, writing, copying, moving, and deleting files, plus running shell commands, each command run as its own process and classified into three permission levels: automatically allowed, requiring user confirmation, or blocked outright. The project is written in TypeScript and Node.js, using LangGraph and LangChain for the agent logic, React and Ink for the terminal interface, and Fastify with WebSockets for the web interface. It is installed as a global npm package and can be run from the terminal or launched with a web flag to open a local browser interface. The project is licensed under the Apache License 2.0.