Tau is a terminal-based coding agent built by Hugging Face that lets you interact with an AI through text commands like "explain this repo" or "fix this stack trace." The agent can read files, edit code, run shell commands, and maintain session history across interactions. What sets it apart is that it's designed to be readable—the codebase is intentionally structured as a teaching project so you can understand how a coding agent actually works without wading through a massive production system. The architecture splits cleanly into three layers: tau_ai handles model provider abstraction, tau_agent contains the core logic (messages, tools, event loop, session management), and tau_coding wraps everything as a usable CLI and terminal UI. The boundary between the reusable brain and the coding-specific environment is deliberate, so the core logic doesn't depend on Textual, Rich, configuration paths, or any particular frontend. Installation is straightforward—you can grab it from PyPI with `uv tool install tau-ai`, pipx, conda-forge, or one of the shell installers for macOS/Linux and Windows. It requires Python 3.12 or newer. Once installed, you run `tau` from your project directory and start typing requests. The agent supports OpenAI, Anthropic, OpenRouter, Hugging Face, and custom endpoints. Sessions are stored as append-only JSONL files under `~/.tau/sessions/`, which means you can resume work, branch conversations, and inspect the full history. The tool set includes read, write, edit, and bash commands, plus slash commands for login, model selection, and session management. The design philosophy emphasizes simplicity and composability: each layer does one job and can be understood independently, tools are ordinary typed functions with schemas, and everything communicates through a typed event stream so you can build custom frontends on top of the same core harness. The documentation includes a build journal in dev-notes/ that tracks the implementation phase by phase, not just the final result. Development happens openly on GitHub under MIT license, with the roadmap tracked in issue #1.