Steward：极简高精度终端 AI 编程助手

The minimal, high-precision AI engineering assistant for your terminal.

Steward is a plug-and-play terminal pair programmer built with a focused philosophy: maximum engineering leverage with zero bloat.

Instead of heavy framework abstractions or sprawling plugin layers, Steward delivers a clean, cohesive tool from day one—giving you the speed of a lightweight coding agent with the safety, resilience, and depth required for serious software development.

┌─────────────────────────────────────────────────────────────┐
│ ✦ Fast, zero-lag Alternate-Screen TUI (Synchronized Mode 2026)│
│ ✦ Unified streaming AI engine with automatic retry          │
│ ✦ Ultra-lean system prompt (~450 tokens) & lazy skill loading│
│ ✦ 16 native tools: search, read, edit, bash, tasks, fetch   │
│ ✦ Content-addressed checkpoints & atomic rewind undo        │
│ ✦ 11 providers: Anthropic, OpenAI, Gemini, OpenRouter, etc. │
│ ✦ Plug-and-play: browser OAuth or standard environment keys │
└─────────────────────────────────────────────────────────────┘

Quick Start

Install Steward globally with your favorite package manager or one-line installer:

One-Line Install

# macOS, Linux & Termux
curl -fsSL https://raw.githubusercontent.com/sapirrior/steward/main/installer/install.sh | bash

# Windows (PowerShell)
irm https://raw.githubusercontent.com/sapirrior/steward/main/installer/install.ps1 | iex

Via Package Managers

npm install -g steward-cli
# or
bun add -g steward-cli

Launch

Navigate to any codebase and launch:

steward

Connect your provider in seconds using /login (Anthropic, OpenRouter, GitHub Copilot) or by setting standard environment keys (OPENAI_API_KEY, GEMINI_API_KEY, ANTHROPIC_API_KEY, etc.).

Why Steward?

🎯 Pure & Focused from Day One

Steward is engineered to be clean and immediately productive without setup friction. You don't need to configure complex plugin trees, manage external SDK wrappers, or debug fragile hook pipelines. Everything you need to inspect, edit, test, and ship code is built directly into the core runtime.

🪶 Ultra-Lean System Prompt & Smart Lazy Loading

Rather than burning thousands of tokens on monolithic boilerplate prompts or dumping entire skill libraries into upfront context, Steward uses a razor-sharp base system prompt (~450 tokens). Context rules and domain capabilities are discovered and loaded on demand through native catalog tools (skill_list, skill_read), preserving maximum model attention and context window budget for your code.

🛡️ Safe by Default, Fearless by Design

• Permission Gated: Destructive file modifications and shell commands always ask for your explicit approval with line-by-line diff previews.
• Atomic Rewind Checkpoints: Every file edit automatically snapshots a pre-mutation image. Made a mistake or want to revert an approach? Use /rewind to cleanly roll back your workspace to any previous turn.

⚡ Multi-Provider AI Core

Steward delivers a unified inference and tool-calling engine powered by AI SDK v7, with resilient rate-limit recovery (with 10-attempt backoff), dynamic model discovery, thinking/reasoning control, and offline resilience across 11 providers.

🎨 Native Terminal UI (stitchable)

Built on a custom line-differential rendering engine with Mode 2026 synchronized output. Enjoy flicker-free markdown streaming, real-time tool badges, fuzzy file searching (@), and interactive slash palettes (/).

Supported Providers

Steward works with any model you prefer—frontier cloud models, unified gateways, or local offline servers:

Keyboard Shortcuts & Slash Commands

Contributing

We welcome contributions! Check out CONTRIBUTING.md for architecture guidelines, development setup, and testing instructions.

License

MIT © sapirrior