Kode-CLI：面向后人类工作流的AI编程代理工具

Core Overview Kode-CLI is a terminal AI programming assistant designed for post-human workflows, developed/maintained by shareAI-lab and hosted on GitHub (released 2026-06-05). It supports AGENTS.md standard, multi-model collaboration, sub-agent systems, and intelligent code completion, providing a unified interface for code understanding, editing, command execution, and workflow automation.

Key Keywords: AI编程, terminal tool, AGENTS.md, multi-model collaboration, sub-agent, code completion, workflow automation, TypeScript, CLI tool, intelligent agent.

Problem Context Existing AI programming tools face limitations: single-model dependency, lack of standardized project context transfer, and suboptimal permission management.

Kode-CLI's Positioning Positioned as an 'AI programming agent for post-human workflows', it assigns dedicated AI unit agents to each task. Built with TypeScript, distributed via npm, and supports Windows native binary execution (no WSL/Git Bash required).

AGENTS.md Support

• Native support for OpenAI's AGENTS.md standard (used by 60k+ projects) for context transfer.
• Instruction discovery: hierarchical search (root to current dir) for AGENTS.override.md/AGENTS.md, backward compatible with CLAUDE.md.
• File拼接 (max 32KB, adjustable via KODE_PROJECT_DOC_MAX_BYTES env var).

Multi-Model Collaboration

• Model咨询: @ask-model-name syntax (e.g., @ask-claude-sonnet-4).
• Sub-agent delegation: @run-agent-name for task orchestration.
• Shortcut: Alt+M to switch models.

Code & Workflow Features

• Code understanding: project structure analysis, smart editing (with human review in safe mode).
• Command execution: sandboxed shell commands, workflow automation via natural language (e.g., "add Docker support").
• ACP protocol support for integration with Toad/Zed.

Smart Fuzzy Matching

• Hyphen-aware (e.g., 'dao' matches 'run-agent-dao-qi-harmony-designer').
• Abbreviation support (e.g., 'dq' → 'dao-qi', 'nde' → 'node').
• Number suffix handling (e.g., 'py3' → 'python3').
• Multi-algorithm fusion (7+ methods) for sorted results.

Context-Aware Completion

• No @ prefix needed (e.g., 'gp5' → '@ask-gpt-5').
• Auto @ prefix on Tab/Enter.
• Mixed completion (commands/files/agents/models).

Unix Command Optimization

• Shows only PATH-existing commands.
• Prioritizes common commands (git, npm, docker).
• Dynamic PATH loading for custom commands.

Permission Modes

• Default YOLO mode: bypasses all checks (max productivity, for trusted environments).
• Safe mode: kode --safe enables permission checks and manual approval.

Model Suggestions

• Recommend task-completion models (avoid GPT-4o/Gemini 2.5 Pro, which are Q&A-optimized).

Privacy

• No default telemetry.
• Explicit network requests only: model provider calls, web tools, plugin downloads, optional updates.

Installation

• npm: npm install -g @shareai-lab/kode (China mirror: add --registry=https://registry.npmmirror.com; dev channel: @dev).
• Aliases: kode, kwa, kd.
• Windows native binary: auto-downloaded post-install (fallback to Node.js if needed; config via env vars like KODE_BINARY_BASE_URL).

Usage Modes

• Interactive: kode (natural language对话).
• Non-interactive: kode -p "query" path/to/file.
• Editor integration: Alt+G to open in $EDITOR; Alt+Enter for newlines.

Evolution Direction Kode-CLI represents the shift from single-function tools to comprehensive agents, closed to open standards, simple Q&A to workflow orchestration.

Core Advantages

1. Standardized context via AGENTS.md.
2. Multi-model flexibility (no vendor lock-in).
3. Sub-agent architecture for complex tasks.
4. Native binary performance.
5. Safety balance (YOLO/safe modes).

Outlook As AGENTS.md and multi-model collaboration mature, Kode-CLI will play an increasingly important role in dev workflows.