中文文档 | Contributing | Documentation

2025-12-22: npm + optionalDependencies distribution (all platforms). Kode prefers the per-platform native binary package (@shareai-lab/kode-bin-*) with Node.js as a fallback; standalone binaries are also published on GitHub Releases. See docs/binary-distribution.md.

Kode supports the AGENTS.md standard: a simple, open format for guiding coding agents, used by 60k+ open-source projects.

• ✅ AGENTS.md - Native support for the OpenAI-initiated standard format
• ✅ Legacy .claude compatibility - Reads .claude directories and CLAUDE.md when present (see docs/compatibility.md)
• ✅ Subagent System - Advanced agent delegation and task orchestration
• ✅ Cross-platform - Works with 20+ AI models and providers

Use # Your documentation request to generate and maintain your AGENTS.md file automatically, while preserving compatibility with existing .claude workflows.

• Kode reads project instructions by walking from the Git repo root → current working directory.
• In each directory, it prefers AGENTS.override.md over AGENTS.md (at most one file per directory).
• Discovered files are concatenated root → leaf (combined size capped at 32 KiB by default; override with KODE_PROJECT_DOC_MAX_BYTES).
• If CLAUDE.md exists in the current directory, Kode also reads it as a legacy instruction file.

Kode is a powerful AI assistant that lives in your terminal. It can understand your codebase, edit files, run commands, and handle entire workflows for you.

⚠️ Security Notice: Kode runs in YOLO mode by default (equivalent to the --dangerously-skip-permissions flag), bypassing all permission checks for maximum productivity. YOLO mode is recommended only for trusted, secure environments when working on non-critical projects. If you're working with important files or using models of questionable capability, we strongly recommend using kode --safe to enable permission checks and manual approval for all operations.

📊 Model Performance: For optimal performance, we recommend using newer, more capable models designed for autonomous task completion. Avoid older Q&A-focused models like GPT-4o or Gemini 2.5 Pro, which are optimized for answering questions rather than sustained independent task execution. Choose models specifically trained for agentic workflows and extended reasoning capabilities.

• Kode does not send product telemetry/analytics by default.
• Network requests happen only when you explicitly use networked features:
  • Model provider requests (Anthropic/OpenAI-compatible endpoints you configure)
  • Web tools (WebFetch, WebSearch)
  • Plugin marketplace downloads (GitHub/URL sources) and OAuth flows (when used)
  • Optional update checks (opt-in via autoUpdaterStatus: enabled)

• 🤖 AI-Powered Assistance - Uses advanced AI models to understand and respond to your requests
• 🔄 Multi-Model Collaboration - Flexibly switch and combine multiple AI models to leverage their unique strengths
• 🦜 Expert Model Consultation - Use @ask-model-name to consult specific AI models for specialized analysis
• 👤 Intelligent Agent System - Use @run-agent-name to delegate tasks to specialized subagents
• 📝 Code Editing - Directly edit files with intelligent suggestions and improvements
• 🔍 Codebase Understanding - Analyzes your project structure and code relationships
• 🚀 Command Execution - Run shell commands and see results in real-time
• 🛠️ Workflow Automation - Handle complex development tasks with simple prompts

• Option+G (Alt+G) opens your message in your preferred editor (respects $EDITOR/$VISUAL; falls back to code/nano/vim/notepad) and returns the text to the prompt when you close it.
• Option+Enter inserts a newline inside the prompt without sending; plain Enter submits. Option+M cycles the active model.

The prompt completes slash commands, file paths, agents, configured models, and available shell commands. Fuzzy and abbreviation matching help surface relevant entries without requiring an exact prefix; accepting an agent or model result adds the required @ prefix.

• 🎨 Interactive UI - Beautiful terminal interface with syntax highlighting
• 🔌 Tool System - Extensible architecture with specialized tools for different tasks
• 💾 Context Management - Smart context handling to maintain conversation continuity
• 📋 Repository Instructions - Reads AGENTS.md and compatible instruction files when present

npm install -g @shareai-lab/kode

🇨🇳 For users in China: If you encounter network issues, use a mirror registry:

npm install -g @shareai-lab/kode --registry=https://registry.npmmirror.com

Kode uses ripgrep (rg) for fast search. By default it is installed via per-platform optionalDependencies (@shareai-lab/kode-ripgrep-<platform>-<arch>). If you install with --no-optional, install system rg or set KODE_RIPGREP_PATH.

Kode also ships an optional per-platform native CLI binary via optionalDependencies (@shareai-lab/kode-bin-<platform>-<arch>). If you install with --no-optional / --omit=optional, it runs via the Node.js entry (dist/index.js).

Kode does not download anything from GitHub during install. (The optional standalone binaries live on GitHub Releases, separate from npm.)

Dev channel (latest features):

npm install -g @shareai-lab/kode@dev

After installation, you can use any of these commands:

• kode - Primary command
• kwa - Kode With Agent (alternative)
• kd - Ultra-short alias

For users who prefer a “portable” executable (no npm install), download the Bun-compiled asset from GitHub Releases:

• https://github.com/shareAI-lab/Kode-CLI/releases

See docs/binary-distribution.md for details (asset names, local build).

• Global config (models, pointers, theme, etc): ~/.kode.json (or <KODE_CONFIG_DIR>/config.json when KODE_CONFIG_DIR is set).
• Project/local settings (output style, etc): ./.kode/settings.json and ./.kode/settings.local.json (legacy .claude is supported for some features).
• Configure models via /model (UI) or kode models import/export (YAML). Details: docs/develop/configuration.md.

Start an interactive session:

kode
# or
kwa
# or
kd

Get a quick response:

kode -p "explain this function" path/to/file.js
# or
kwa -p "explain this function" path/to/file.js

Run Kode as an ACP agent server (stdio JSON-RPC), for clients like Toad/Zed:

kode-acp
# or
kode --acp

Toad example:

toad acp "kode-acp"

More: docs/acp.md.

Kode supports a powerful @ mention system for intelligent completions:

# Consult specific AI models for expert opinions
@ask-claude-sonnet-4 How should I optimize this React component for performance?
@ask-gpt-5 What are the security implications of this authentication method?
@ask-o1-preview Analyze the complexity of this algorithm

# Delegate tasks to specialized subagents
@run-agent-simplicity-auditor Review this code for over-engineering
@run-agent-architect Design a microservices architecture for this system
@run-agent-test-writer Create comprehensive tests for these modules

# Reference files and directories with auto-completion
@packages/core/src/query/index.ts
@docs/README.md
@.env.example

The @ mention system provides intelligent completions as you type, showing available models, agents, and files.

Kode can connect to MCP servers to extend tools and context.

• Config files: .mcp.json (recommended) or .mcprc in your project root. See docs/mcp.md.
• CLI:

kode mcp add
kode mcp list
kode mcp get <name>
kode mcp remove <name>

Example .mcprc:

{
  "my-sse-server": { "type": "sse", "url": "http://127.0.0.1:3333/sse" }
}

• Default mode skips most prompts for speed.
• Safe mode: kode --safe requires approval for Bash commands and file writes/edits.
• Plan mode: the assistant may ask to enter plan mode to draft a plan file; while in plan mode, only read-only/planning tools (and the plan file) are allowed until you approve exiting plan mode.

• Multi-line/large paste is inserted as a placeholder and expanded on submit.
• Pasting multiple existing file paths inserts @path mentions automatically (quoted when needed).
• Image paste (macOS): press Ctrl+V to attach clipboard images; you can paste multiple images before sending.

• In safe mode (or with KODE_SYSTEM_SANDBOX=1), agent-triggered Bash tool calls try to run inside a bwrap sandbox when available.
• Network is disabled by default; set KODE_SYSTEM_SANDBOX_NETWORK=inherit to allow network.
• Set KODE_SYSTEM_SANDBOX=required to fail closed if sandbox cannot be started.
• See docs/system-sandbox.md for details and platform notes.

• Models: use /model, or kode models import kode-models.yaml, and ensure required API key env vars exist.
• Windows: install via npm as usual; if you want a single-file executable, use the GitHub Release asset for your platform.
• MCP: use kode mcp list to check server status; tune MCP_CONNECTION_TIMEOUT_MS, MCP_SERVER_CONNECTION_BATCH_SIZE, and MCP_TOOL_TIMEOUT if servers are slow.
• Sandbox: install bwrap (bubblewrap) on Linux, or set KODE_SYSTEM_SANDBOX=0 to disable.

Use the # prefix to generate and maintain your AGENTS.md documentation:

# Generate setup instructions
# How do I set up the development environment?

# Create testing documentation
# What are the testing procedures for this project?

# Document deployment process
# Explain the deployment pipeline and requirements

This mode automatically formats responses as structured documentation and appends them to your AGENTS.md file.

# Clone the repository
git clone https://github.com/shareAI-lab/Kode-CLI.git
cd Kode

# Build the image locally
docker build --no-cache -t kode .

# Run in your project directory
cd your-project
docker run -it --rm \
  -v "$PWD:/workspace" \
  -v "$HOME/.kode:/home/node/.kode" \
  -v "$HOME/.kode.json:/home/node/.kode.json" \
  -w /workspace \
  kode

The Docker setup includes:

• Volume Mounts:

  • $PWD:/workspace - Mounts your current project directory
  • $HOME/.kode:/home/node/.kode - Preserves your Kode data between runs
  • $HOME/.kode.json:/home/node/.kode.json - Preserves your global configuration

• Working Directory: Set to /workspace inside the container

• Interactive Mode: Uses -it flags for interactive terminal access

• Cleanup: --rm flag removes the container after exit

Note: The image runs as the unprivileged node user. Create the two host configuration paths before mounting them if they do not already exist.

The first time you run the Docker command, it will build the image. Subsequent runs will use the cached image for faster startup.

You can use the onboarding to set up the model, or /model. If you don't see the models you want on the list, you can manually set them in /config As long as you have an openai-like endpoint, it should work.

• /help - Show available commands
• /model - Change AI model settings
• /config - Open configuration panel
• /agents - Manage subagents
• /output-style - Set the output style
• /statusline - Configure a custom status line command
• /cost - Show token usage and costs
• /clear - Clear conversation history
• /init - Initialize project context
• /plugin - Manage plugins/marketplaces (skills, commands)

Kode supports subagents (agent templates) for delegation and task orchestration.

• Agents are loaded from .kode/agents and .claude/agents (user + project), plus plugins/policy and --agents.
• Manage in the UI: /agents (creates new agents under ./.kode/agents / ~/.kode/agents by default; legacy .claude/agents is read-compat).
• Run via mentions: @run-agent-<agentType> ...
• Run via tooling: Task(subagent_type: "<agentType>", ...)
• CLI flags: --agents <json> (inject agents for this run), --setting-sources user,project,local (control which sources are loaded)

Minimal agent file example (./.kode/agents/reviewer.md):

---
name: reviewer
description: 'Review diffs for correctness, security, and simplicity'
tools: ['Read', 'Grep']
model: inherit
---

Be strict. Point out bugs and risky changes. Prefer small, targeted fixes.

Model field notes:

• Compatibility aliases: inherit, opus, sonnet, haiku (mapped to model pointers)
• Kode selectors (via /model): pointers (main|task|compact|quick), profile name, modelName, or provider:modelName (e.g. openai:o3)

Validate agent templates:

kode agents validate

See docs/agents-system.md.

Kode supports:

• Agent Skills format (SKILL.md) for reusable skill packs
• Marketplace compatibility (.kode-plugin/marketplace.json, legacy .claude-plugin/marketplace.json) for sharing/installing skill packs

# Add a marketplace (local path, GitHub owner/repo, or URL)
kode plugin marketplace add ./path/to/marketplace-repo
kode plugin marketplace add owner/repo
kode plugin marketplace list

# Install a plugin pack (installs skills/commands)
kode plugin install document-skills@anthropic-agent-skills --scope user

# Project-scoped install (writes to ./.kode/...)
kode plugin install document-skills@anthropic-agent-skills --scope project

# Disable/enable an installed plugin
kode plugin disable document-skills@anthropic-agent-skills --scope user
kode plugin enable document-skills@anthropic-agent-skills --scope user

Interactive equivalents:

/plugin marketplace add owner/repo
/plugin install document-skills@anthropic-agent-skills --scope user

• In interactive mode, run a skill as a slash command: /pdf, /xlsx, etc.
• Kode can also invoke skills automatically via the Skill tool when relevant.

Create ./.kode/skills/<skill-name>/SKILL.md (project) or ~/.kode/skills/<skill-name>/SKILL.md (user):

---
name: my-skill
description: Describe what this skill does and when to use it.
allowed-tools: Read Bash(git:*) Bash(jq:*)
---

# Skill instructions

Naming rules:

• name must match the folder name
• Lowercase letters/numbers/hyphens only, 1–64 chars

Compatibility:

• Kode also discovers .claude/skills and .claude/commands for legacy compatibility.

• Marketplace repo: publish a repo containing .kode-plugin/marketplace.json listing plugin packs and their skills directories (legacy .claude-plugin/marketplace.json is also supported).
• Plugin repo: for full plugins (beyond skills), include .kode-plugin/plugin.json at the plugin root and keep all paths relative (./...).

See docs/skills.md for a compact reference and examples.

Use output styles to switch system-prompt behavior.

• Select: /output-style (menu) or /output-style <style>
• Built-ins: default, Explanatory, Learning
• Stored per-project in ./.kode/settings.local.json as outputStyle (legacy .claude/settings.local.json is supported)
• Custom styles: Markdown files under output-styles/ in .claude/.kode user + project locations
• Plugins can provide styles under output-styles/ (or manifest outputStyles); plugin styles are namespaced as <plugin>:<style>

See docs/output-styles.md.

Unlike single-model CLIs, Kode implements true multi-model collaboration, allowing you to fully leverage the unique strengths of different AI models.

We designed a unified ModelManager system that supports:

• Model Profiles: Each model has an independent configuration file containing API endpoints, authentication, context window size, cost parameters, etc.
• Model Pointers: Users can configure default models for different purposes in the /model command:
  • main: Default model for main Agent
  • task: Default model for SubAgent
  • compact: Model used for automatic context compression when nearing the context window
  • quick: Fast model for simple operations and utilities
• Dynamic Model Switching: Support runtime model switching without restarting sessions, maintaining context continuity

You can export/import model profiles + pointers as a team-shareable YAML file. By default, exports do not include plaintext API keys (use env vars instead).

# Export to a file (or omit --output to print to stdout)
kode models export --output kode-models.yaml

# Import (merge by default)
kode models import kode-models.yaml

# Replace existing profiles instead of merging
kode models import --replace kode-models.yaml

# List configured profiles + pointers
kode models list

Example kode-models.yaml:

version: 1
profiles:
  - name: OpenAI Main
    provider: openai
    modelName: gpt-4o
    maxTokens: 8192
    contextLength: 128000
    apiKey:
      fromEnv: OPENAI_API_KEY
pointers:
  main: gpt-4o
  task: gpt-4o
  compact: gpt-4o
  quick: gpt-4o

Our specially designed TaskTool (Architect tool) implements:

• Subagent Mechanism: Can launch multiple sub-agents to process tasks in parallel
• Model Parameter Passing: Users can specify which model SubAgents should use in their requests
• Default Model Configuration: SubAgents use the model configured by the task pointer by default

We specially designed the AskExpertModel tool:

• Expert Model Invocation: Allows temporarily calling specific expert models to solve difficult problems during conversations
• Model Isolation Execution: Expert model responses are processed independently without affecting the main conversation flow
• Knowledge Integration: Integrates expert model insights into the current task

• Option+M Quick Switch: Press Option+M in the input box to cycle the main conversation model
• /model Command: Use /model command to configure and manage multiple model profiles, set default models for different purposes
• User Control: Users can specify specific models for task processing at any time

Architecture Design Phase

• Use o3 model or GPT-5 model to explore system architecture and formulate sharp and clear technical solutions
• These models excel in abstract thinking and system design

Solution Refinement Phase

• Use gemini model to deeply explore production environment design details
• Leverage its deep accumulation in practical engineering and balanced reasoning capabilities

Code Implementation Phase

• Use Qwen Coder model, Kimi k2 model, GLM-4.5 model, or Claude Sonnet 4 model for specific code writing
• These models have strong performance in code generation, file editing, and engineering implementation
• Support parallel processing of multiple coding tasks through subagents

Problem Solving

• When encountering complex problems, consult expert models like o3 model, Claude Opus 4.1 model, or Grok 4 model
• Obtain deep technical insights and innovative solutions

# Example 1: Architecture Design
"Use o3 model to help me design a high-concurrency message queue system architecture"

# Example 2: Multi-Model Collaboration
"First use GPT-5 model to analyze the root cause of this performance issue, then use Claude Sonnet 4 model to write optimization code"

# Example 3: Parallel Task Processing
"Use Qwen Coder model as subagent to refactor these three modules simultaneously"

# Example 4: Expert Consultation
"This memory leak issue is tricky, ask Claude Opus 4.1 model separately for solutions"

# Example 5: Code Review
"Have Kimi k2 model review the code quality of this PR"

# Example 6: Complex Reasoning
"Use Grok 4 model to help me derive the time complexity of this algorithm"

# Example 7: Solution Design
"Have GLM-4.5 model design a microservice decomposition plan"

// Example of multi-model configuration support
{
  "modelProfiles": [
    { "name": "o3", "provider": "openai", "modelName": "o3", "apiKey": "...", "maxTokens": 1024, "contextLength": 128000, "isActive": true, "createdAt": 1710000000000 },
    { "name": "qwen", "provider": "alibaba", "modelName": "qwen-coder", "apiKey": "...", "maxTokens": 1024, "contextLength": 128000, "isActive": true, "createdAt": 1710000000001 }
  ],
  "modelPointers": {
    "main": "o3",           // Main conversation model
    "task": "qwen-coder",   // Sub-agent model
    "compact": "o3",        // Context compression model
    "quick": "o3"           // Quick operations model
  }
}

• Usage Statistics: Use /cost command to view token usage and costs for each model
• Multi-Model Cost Comparison: Track usage costs of different models in real-time
• History Records: Save cost data for each session

• Context Inheritance: Maintain conversation continuity when switching models
• Context Window Adaptation: Automatically adjust based on different models' context window sizes
• Session State Preservation: Ensure information consistency during multi-model collaboration

1. Maximized Efficiency: Each task is handled by the most suitable model
2. Cost Optimization: Use lightweight models for simple tasks, powerful models for complex tasks
3. Parallel Processing: Multiple models can work on different subtasks simultaneously
4. Flexible Switching: Switch models based on task requirements without restarting sessions
5. Leveraging Strengths: Combine advantages of different models for optimal overall results

This multi-model collaboration capability makes Kode a true AI Development Workbench, not just a single AI assistant.

Kode is built with modern tools and requires Bun for development.

# macOS/Linux
curl -fsSL https://bun.sh/install | bash

# Windows
powershell -c "irm bun.sh/install.ps1 | iex"

# Clone the repository
git clone https://github.com/shareAI-lab/Kode-CLI.git
cd kode

# Install dependencies
bun install --frozen-lockfile

# Run in development mode
bun run dev

bun run build

# Run tests
bun test

# Test the CLI
./cli.js --help

We welcome contributions! Please see our Contributing Guide for details.

Apache 2.0 License - see LICENSE for details.

• Some code from @dnakov's anonkode
• Some UI learned from gemini-cli
• Some system design learned from upstream agent CLIs

• 📚 Documentation
• 🐛 Report Issues
• 💬 Discussions