What is Agent CLI?

Agent CLI is a local-first AI Agent runtime workspace. It transforms your development intent into recoverable Sessions, observable Runs, auditable Traces, and project memory — all stored within your engineering workspace.

It is not a chatbot wrapped in a terminal. Alius is built for software that can participate in its own evolution: understanding goals, generating plans, executing through controlled runtime boundaries, recording evidence, reviewing results, and carrying learnings into the next iteration.

Core Features

  • Plan-driven workspace — Not a simple chat interface, but an Agent runtime built around goal understanding, plan generation, controlled execution, and result review
  • Dual-mode execution — Plan mode for complex multi-step tasks, Bypass mode for quick single-turn interactions
  • Project-level configuration — Model pool, role identity, and security policies managed per project
  • Layered memory — Working memory, episodic memory, and semantic memory accumulate project knowledge across Sessions
  • Security boundary — Shell Gate performs risk analysis and approval control on all shell and file operations
  • Local-first — All data stored in the .alius/ directory, no data upload required

System Requirements

  • macOS 12+ or Linux (x86_64 / ARM64)
  • Windows 10+ (PowerShell)
  • Node.js 18+ (for npm installation)
  • Rust 1.75+ (for building from source only)

Quick Install

curl -fsSL https://alius.tech/install | sh

Or install via a package manager:

# npm
npm install -g @alius-tech/alius

# Homebrew
brew tap AliusTech/tap
brew install alius

Verify the installation:

alius --version

Initialize a Project

Run in your project directory:

cd your-project
alius init

alius init launches an interactive initialization wizard that guides you through:

  1. Language selection — Choose the workspace interface language
  2. Model pool configuration — Import models from supported providers (BigModel, Xiaomi MiMo, DeepSeek, etc.)
  3. Model assignment — Assign models to the Plan, Execute, and Review roles
  4. Soul selection — Choose the Agent’s role identity

After initialization, a .alius/ structure is generated in your project directory:

your-project/
├── .alius/
│   ├── config/         # Project configuration (providers.toml, model.toml, soul.toml, etc.)
│   ├── memory/         # Project memory storage
│   └── workspace/      # Workspace documents
├── src/
└── ...

The initialization wizard is resumable. If you interrupt it midway, running alius init again lets you continue from where you left off.

Enter the Workspace

Launch the interactive workspace:

alius

You will see the Ratatui-based TUI workspace interface, containing:

  • Top bar — Version, Soul, mode, and network status
  • Session area — Workflow conversation blocks (Request, Understanding, Plan Proposal, Execution, Prompt, Output, Error)
  • Plan panel — Plan nodes, status, acceptance criteria, and evidence (appears only after Plan mode approval)
  • Interaction area — Text input, option selection, and approval controls
  • Status bar — Current workspace, Git status, and runtime status

Two Execution Modes

In the TUI workspace, you can switch between two modes using Shift+Tab:

For complex multi-step tasks. The workflow:

  1. You describe a goal
  2. The model proactively asks clarifying questions about unclear requirements (as multiple-choice or fill-in-the-blank)
  3. The model generates an executable plan list
  4. You approve the plan
  5. Approved plan nodes execute step by step
  6. After execution completes, you confirm to close the plan panel

Plan mode supports up to 20 iterations, can invoke tools, execute shell commands, and check for convergence.

Bypass Mode

For quick single-turn interactions — questions, clarifications, lightweight analysis. Bypass mode sends input directly to single-turn execution. The model may request one tool call, but will not enter a multi-turn iteration loop.

One-shot Execution

If you don’t need the interactive workspace, you can run a single prompt directly:

alius run -p "Perform a structured code review on this module"

This is a one-shot command that executes directly in Chat mode and streams the result to standard output.

Next Steps

  • Installation Guide — Detailed installation options, platform support, and global parameters
  • Quick Start Guide — Follow the step-by-step tutorial for your first workflow
  • Core Concepts — Deep dive into workspaces, Sessions, and modes
  • Commands — CLI commands and TUI slash command reference
  • Configuration — Configure providers, model pools, and security policies