Core concepts

Tau is built from a small set of concepts. Understand these and the rest of the docs (and the agent itself) will make sense.

The agent loop

When you send a prompt, Tau runs an agent loop: it asks the model to respond, streams that response, and when the model asks to use a tool, Tau runs the tool, feeds the result back, and asks the model to continue — repeating until the model has nothing left to do. That loop is the heart of every coding agent. → How Tau works: the agent loop

Providers and models

A provider is the service that hosts AI models (OpenAI, Anthropic, OpenAI Codex, OpenRouter, Hugging Face, or any OpenAI-compatible endpoint). A model is the specific brain you’re talking to (e.g. gpt-5.5, claude-sonnet-4-6). You pick a provider + model; you can switch either mid-session. → Providers & models

Tools

Tools are the actions the agent can take in your project. Tau ships four: read, write, edit, and bash. The model decides when to call them; Tau executes them in your working directory and streams the results. → Tools reference

Sessions

A session is one ongoing conversation plus everything the agent did in it. Sessions are saved to disk as append-only files, so you can resume them later. Because the history is a tree, you can also branch from any earlier point to try a different direction, and export a session to HTML or JSONL. → Sessions

Project instructions (AGENTS.md)

An AGENTS.md file gives the agent standing instructions about your project — conventions, gotchas, how to run tests. Tau discovers these files automatically and folds them into the system prompt. → Project instructions

Skills and prompt templates

Skills are Markdown files describing how to do a specific task; you invoke one with /skill:<name>. Prompt templates are reusable prompts you trigger by name (with optional variables) instead of retyping. Both can live at the user level (all projects) or per-project. → Skills & prompt templates

Context and compaction

The model can only “see” a limited amount of text at once — its context window. Long sessions fill it up. Tau estimates usage and compacts automatically: it summarizes older messages so the conversation can keep going. You can also compact on demand with /compact. → Managing context

Thinking modes

Some models can spend extra effort “thinking” before they answer. Tau exposes a thinking mode (off → minimal → low → medium → high → xhigh) you can cycle when the active model supports it, and optionally show the streamed reasoning. → Managing context

Two interfaces

Tau runs as an interactive TUI (the default) or in print mode (-p) for a single non-interactive prompt. Both share the same session environment. → The interactive session · Print mode