tau: Python Terminal Coding Agent Inspired by Pi

Overview

Tau is a small, readable terminal coding agent, built as a Python port of Pi’s minimalist coding agent. It helps developers ask for work in a project directory, such as explaining a codebase or fixing an error, then lets the agent read and edit files and run shell commands. Tau is also designed as a teaching project for people who want to understand how coding agents are assembled.

Its Python packages separate provider streaming, a reusable agent harness, and the coding environment. A terminal UI and print mode are included, while the harness emits events so other frontends can consume the same agent behavior. Tau supports Python 3.12 and newer; its README describes provider integrations and custom OpenAI-compatible endpoints, rather than requiring one particular model provider.

Key Features

• Interactive terminal UI and non-interactive print mode.
• Built-in read, write, edit, and bash tools for working in a project.
• Append-only JSONL session history with resume and branching.
• Project instructions and resources loaded from AGENTS.md, .tau/, and .agents/.
• User skills, prompt templates, slash commands, and custom TUI themes.
• Manual and optional automatic context compaction.
• Provider-neutral events that can drive the included UI, text output, JSON, transcripts, or custom frontends.

Use Cases

• Developers who want a terminal agent to inspect a repository, explain its structure, or make code changes.
• Programmers who want to experiment with agent tools, event streams, sessions, and harness design in a relatively small codebase.
• Teams building a custom frontend that want to reuse an agent harness without coupling it to Tau’s Textual UI.
• Users who need resumable coding sessions and project-specific instructions while working across coding tasks.

Getting Started

Install with uv tool install tau-ai, then run tau from the project directory. Tau needs a model provider; connect one with /login. See the README for installer options, provider configuration, and development instructions.

Alternatives

• qwen-code: Unlike tau’s small, teaching-oriented Python core, qwen-code offers multiple interfaces, model APIs, and SDK and scripting options.
• fx: fx is a native Zig agent with an interactive shell and embedding options, while tau is a Python teaching project with a reusable event-driven core.
• zero: Zero is a Go terminal agent emphasizing model choice, scripting, and explicit side-effect controls; tau focuses on a readable Python implementation.
• openinterpreter: Open Interpreter is a Rust agent built around low-cost models and multiple harnesses, while tau is a minimal Python port of Pi.

Considerations

The project is new and changing quickly, with frequent commits and releases. Although it has multiple contributors, contribution activity is highly concentrated with one person. Recent issue closures and pull request merges have moved quickly, but that does not remove the adoption risk of a young codebase. Tau requires Python 3.12 or newer and a configured model provider. The README identifies it as under active development and tracks its roadmap in GitHub issue #1. It is released under the MIT License.