tau

A minimal Rust coding agent harness

Architecture

ai             LLM streaming primitives (providers, models, event streams)
agent          generic agent loop (tools, steering, follow-ups, events)
coding-agent   built-in tools, system prompt, REPL + headless CLI

ai and agent have no opinion about coding — they're generic building blocks. coding-agent is one harness built on top; you could build a data-analysis agent or a research agent on the same foundation without importing coding-specific dependencies.

Tools

Tool	What it does
bash	Shell execution with timeout, cancellation, output truncation
file_read	Read files with line numbers, offset/limit, binary detection
file_write	Write/create files, auto-create parent directories
file_edit	Exact-match string replacement with context on miss
grep	Ripgrep-backed content search with glob filtering
glob	Native gitignore-aware file discovery, mtime-sorted

Hashline mode — tau also ships hash-anchored line editing (invented by Can Boluk for oh-my-pi). Each line gets a content-hash tag; edits reference tags instead of reproducing text.

Installation

From source (recommended for development)

Requires Rust toolchain (1.75+).

# Clone and install
git clone https://github.com/tnguyen21/tau.git
cd tau
cargo install --path coding-agent

# Or install directly from GitHub without cloning
cargo install --git https://github.com/tnguyen21/tau.git coding-agent

This puts tau on your $PATH.

Prebuilt Linux binary

CI builds a static x86_64-unknown-linux-musl binary on every tagged release. Download it directly:

curl -fsSL \
  https://github.com/tnguyen21/tau/releases/latest/download/tau-x86_64-unknown-linux-musl \
  -o /usr/local/bin/tau
chmod +x /usr/local/bin/tau

Override the release source with TAU_BINARY_VERSION, TAU_BINARY_REPO, or TAU_BINARY_URL when needed. See Release and container install for details.

Structural analysis tools (optional)

For call-graph and CFG analysis, install the companion binaries:

cargo install --git https://github.com/tnguyen21/pycg-rs.git pycg
cargo install --git https://github.com/tnguyen21/pycfg-rs.git pycfg

These are only needed if you pass structural tools (cg_*, cfg_*) via --tools.

Quick start

# Set a provider key
export ANTHROPIC_API_KEY=sk-ant-...
# or
export OPENAI_API_KEY=sk-...

# Interactive REPL
tau

# Choose a model
tau --model claude-sonnet-4-6

# Headless (for scripting / benchmarks)
tau --prompt "List all Rust files in this repo"

# Restrict tool access
tau --tools file_read,file_write,file_edit,glob,grep,bash

# With stats
tau --stats --prompt "Explain this repo"

Providers

tau supports three backend surfaces:

• OpenAI Responses API for direct OpenAI models
• Anthropic Messages API for direct Anthropic models
• OpenAI-compatible Chat Completions endpoints, including OpenRouter model families such as Gemini, Qwen, Grok, DeepSeek, and Kimi

There is no separate Kimi provider in tau; Kimi models are accessed through OpenRouter/OpenAI-compatible chat routing.

Testing

cargo test                    # 270+ offline tests, no API keys needed
cargo bench                   # criterion benchmarks (SSE parsing, serde, agent construction)

Live provider tests exist behind a double opt-in gate: OPENAI_API_KEY + RUN_LIVE_PROVIDER_TESTS=1.

Configuration

tau reads ~/.tau/config.toml (global) and .tau/config.toml (project-local):

[agent]
model = "claude-sonnet-4-6"
edit_mode = "hashline"    # or "standard"

[agent.thinking]
level = "medium"          # off, low, medium, high

Sessions

Sessions persist as JSONL in ~/.tau/sessions/. Resume with --resume (latest) or --session <id>.

Project structure

ai/
  src/
    types.rs              # ContentBlock, Message, Model, streaming types
    stream.rs             # EventStream (mpsc + oneshot channels)
    providers/            # OpenAI Responses, Anthropic Messages
    models.rs, catalog.rs # Model registry (~65 models)
agent/
  src/
    loop_.rs              # Two-level agent loop (outer: follow-ups, inner: stream → tools → steer)
    agent.rs              # Agent struct (state, queues, cancellation, events)
    types.rs              # AgentTool trait, AgentEvent, AgentLoopConfig
coding-agent/
  src/
    tools/                # bash, file_read, file_write, file_edit, grep, glob, hashline
    system_prompt.rs      # Dynamic prompt built from active tools + cwd
    cli.rs, config.rs     # CLI parsing, TOML config
    session.rs            # JSONL session persistence
    main.rs               # REPL + headless modes

Docs

• Architecture overview — detailed walkthrough of types, patterns, and design decisions
• Release and container install — tag-driven musl releases and container download flow
• Feature comparison — tau vs 6 other harnesses across every dimension
• Toolset tradeoffs — why this toolset, what others chose, and what it means
• Benchmarks landscape — harness-sensitive benchmarks and key numbers
• Harness lit review — people, papers, and open questions in harness engineering

Roadmap

tau is currently a hackable reference. The path forward:

1. Daily driver — compaction, permissions, sub-agents, MCP, skills. See feature-comparison.md for the full gap analysis.
2. RL trajectory generator — the daily driver bootstraps itself, then generates trajectories for co-training research.

The long-term thesis: models and harnesses will be co-developed. An RL post-training step where the model does rollouts with harness trajectories and learns better tool use, context management, and file navigation. tau is built to be instrumentable enough to make that possible.

License

MIT