ZhiXia — One Command for Every AI Agent CLI

ZhiXia Logo

Describe a task once with one set of unified parameters — ZhiXia translates it into the native invocation of your AI CLI and runs it fully unattended in forced non-interactive mode.

The Problem

Every code AI agent CLI has its own argument system, its own interaction model, and its own silent-run configuration. ClaudeCode-CLI, TRAE-CLI, OpenCode-CLI, Codex-CLI, Qoder-CLI, CodeBuddy-CLI — they all differ.

For developers who need automation and batch scheduling, this means:

No single generic command can drive different AI tools
AI capabilities can’t be wired into CI/CD pipelines
Every new tool means re-learning parameters and rewriting scripts
Automation cost grows linearly with the number of tools

The Solution

ZhiXia acts as a lightweight proxy bridge layer that gives every AI Agent CLI a unified input standard. You learn one command; ZhiXia internally translates it into each tool’s native parameters for fully automatic, non-interactive execution.

Core idea: unified abstraction → driver mapping → verbatim passthrough

Upward: clean, consistent task instruction parameters
Downward: adapts to each vendor CLI’s differentiated implementation
Complete passthrough of stdin/stdout and exit codes — transparent behavior

Core Value

What You Get	How It Works
Zero learning cost	One set of command parameters covers all major AI CLIs — no more per-tool documentation
Automation-friendly	Forced non-interactive mode disables all human prompts; embed directly in CI/CD, batch scheduling, and smart pipelines
Out of the box	Single static binary, zero config, zero dependencies — feels like a system tool (`grep`, `cp`)

Key Features

1. Unified Parameter Abstraction

Encapsulates common task description, working directory, model selection, timeout, and read/write permissions into one consistent set of parameters — write once, run everywhere.

2. Multi-Driver Adaptation

Built-in driver mappings for ClaudeCode-CLI, TRAE-CLI, OpenCode-CLI, Codex-CLI, Qoder-CLI, and CodeBuddy-CLI, automatically translated to each tool’s native arguments.

3. Forced Non-Interactive Mode

Automatically sets each CLI’s silent-run environment variables and flags (e.g. CLAUDE_CODE_NON_INTERACTIVE=1, TRAE_HEADLESS=true) so tasks run 100% unattended in automation scenarios.

4. Raw Argument Passthrough

When the unified abstraction can’t cover a tool’s special features, pass native arguments directly with --raw-args — compatible with all edge cases.

5. Standard Stream & Exit-Code Passthrough

stdin/stdout/stderr are passed through completely and exit codes are inherited unchanged, so upstream scheduling systems can reliably judge task results.

Quick Start

Install at least one target AI CLI (ClaudeCode-CLI, TRAE-CLI, OpenCode-CLI, Codex-CLI, Qoder-CLI, or CodeBuddy-CLI)
Download the zhixia-* binary for your platform from the release page
Put it on your $PATH and rename it to zhixia
Run:

zhixia --tool claude --prompt "重构 src/auth 登录逻辑"
zhixia --tool codex --model gpt-4o --prompt "生成单元测试" --work-dir ./src

Download

Platform	Download	Requirements
Windows	Latest Release	Windows 10+ (x64)
macOS	Latest Release	macOS 11+ (Intel & Apple Silicon)
Linux	Latest Release	Ubuntu 20.04+ / Fedora 36+ (x64)

Chinese users can download from Gitee Releases.

Privacy

Fully local — ZhiXia does not collect or upload any user data
No network — all AI tasks run in your local environment with the tool you choose; ZhiXia itself never accesses the network or sends data to any third-party service

Status

v0.1.0 Released — production-ready and actively maintained.