taui

Taui is a customizable agentic coding interface for developers. It runs as a full-screen Textual TUI; the terminal app is the product.

Alpha software: APIs, commands, and behavior may change.

Install

uvx taui

For local development:

uv run taui
uv run taui --version
uv run taui --login
uv run taui -p copilot -m <model>
uv run taui -p codex -m <model>
uv run taui -d /path/to/project
uv run taui --session <session_id>

CLI parsing and app launch live in taui/main.py:29 and taui/main.py:90.

What It Does

• Runs an async agent loop over provider responses, tool calls, and observations: taui/agent/loop.py:93.
• Wires providers, tools, extensions, prompts, store, and loop in one composition root: taui/session.py:139.
• Stores sessions as append-only SQLite event streams in the working directory: taui/store/store.py:97 and taui/store/stream.py:22.
• Renders chat, streaming output, approvals, questions, the file sidebar, and session modal controls in Textual: taui/tui/app.py:206.

Providers

Built-in providers:

Run taui --login to authenticate. Credentials are loaded through taui/llm_provider/config.py:15 and selected by Config.load() at taui/config.py:64.

Commands

Important slash commands are registered in taui/commands/builtins.py:858.

Keys

App-level bindings are defined in TauiApp.BINDINGS at taui/tui/app.py:399. Input-specific bindings are in ChatInput.BINDINGS at taui/tui/widgets/chat_input.py:64.

Configuration

Config fields are defined in taui/config.py:33.

[taui]
provider = "copilot"
model = "claude-sonnet-4.5"
max_turns = 50
verbose_tools = true

[taui.tool_policy]
bash = "confirm"
write = "confirm"
edit = "confirm"

[taui.permission]
read = { "*" = "allow" }
bash = { "git status" = "allow", "*" = "ask" }

Tool policy evaluation is in taui/tools/executor.py:42; pattern permissions are in taui/permissions.py:38.

Extensions And Skills

Extensions are Python files loaded from ~/.taui/extensions/*.py and .taui/extensions/*.py. The register(ctx) context is defined at taui/extensions/__init__.py:66; extension loading starts at taui/extensions/__init__.py:169.

Skills are SKILL.md files discovered by taui/skills/__init__.py:91 and loaded lazily by taui/skills/__init__.py:52.

Install skills from external sources (compatible with vercel-labs/skills) with /skills add <source>, or just paste npx skills add <source> / a bare repo ref into the chat input. Sources may be GitHub shorthand (owner/repo), a full git/GitHub/GitLab URL, a URL pointing at one skill (…/tree/<ref>/<path>), an SSH git URL, or a local path; add -g to install globally. The installer lives in taui/skills/installer.py, and the self-edit agent exposes the same engine as the install_skill tool.

Documentation

• Product and architecture overview: docs/taui.md:1
• Runtime flow: docs/runtime.md:1
• Tools and permissions: docs/tools.md:1, docs/permission-dsl.md:1
• Extensions, hooks, skills, and agents: docs/build-your-harness.md:1, docs/extension-hooks.md:1, docs/agents.md:1
• Providers and auth: docs/providers.md:1
• Context and prompts: docs/context-strategies.md:1, docs/system-prompt.md:1
• Tests and visual harness: docs/testing.md:1

Checks

uv run ruff check .
uv run python -m pytest -q -m tools  # replace tools with the changed domain

Use domain groups for focused checks instead of running everything:

uv run python -m pytest -q -m tools
uv run python -m pytest -q -m agent
uv run python -m pytest -q -m "agent or provider"
uv run python -m pytest -q -m tui

Target focused tests first when changing one subsystem. The scenario and visual harness are documented in docs/testing.md:1.