Skip to content

AGENTS.md

Latest commit

 

History

History
234 lines (185 loc) · 7.03 KB

AGENTS.md

File metadata and controls

234 lines (185 loc) · 7.03 KB

Agents

CodeMie Native (Built-in)

LangGraph-based coding assistant with no installation required.

Features:

  • Modern terminal UI with streaming responses
  • File operations with intelligent filtering
  • Command execution with progress tracking
  • Planning and todo management tools
  • Interactive conversations with context memory
  • Task-focused execution mode
  • Cross-platform support (Windows, Linux, macOS)

Available Tools:

  • read_file - Read file contents with progress tracking for large files
  • write_file - Write content to files with automatic directory creation
  • list_directory - List files and directories with intelligent filtering (auto-filters node_modules, .git, build artifacts, etc.)
  • execute_command - Execute shell commands with security checks and progress estimation
  • write_todos - Create or update structured todo lists for planning
  • update_todo_status - Update status of specific todos (pending, in_progress, completed)
  • append_todo - Add new todo items to existing lists
  • clear_todos - Clear all todos from the list
  • show_todos - Display current todo list with progress information

Security Features:

  • Path traversal prevention (restricted to working directory)
  • Dangerous command blocking (rm -rf, sudo, etc.)
  • Configurable directory and pattern filtering
  • Secure file access controls

Usage:

codemie-code                    # Interactive mode
codemie-code "task"             # Start with message
codemie --task "task"           # Single task execution
codemie-code health             # Health check

Claude Code

Anthropic's official CLI with advanced code understanding.

Installation: codemie install claude

Features:

  • Advanced code understanding and generation
  • Multi-file editing capabilities
  • Project-aware context
  • Interactive conversations
  • Non-interactive mode with -p flag

Usage:

codemie-claude                   # Interactive mode
codemie-claude "message"         # Start with message
codemie-claude -p "message"      # Non-interactive/print mode
codemie-claude health            # Health check

Claude Code ACP

Claude Code adapter for IDE integration via ACP (Agent Communication Protocol).

Installation: codemie install claude-acp

What is ACP? ACP (Agent Communication Protocol) is a stdio-based JSON-RPC protocol that enables AI agents to communicate with editors and IDEs. It provides a standardized way for editors like Zed, JetBrains, and Emacs to integrate with AI coding assistants.

Requirements:

  • Node.js 20.0.0 or higher
  • Supported providers: LiteLLM, AI/Run SSO, or direct Anthropic API access
  • IDE with ACP support (Zed, JetBrains, Emacs, etc.)

Features:

  • Stdio-based JSON-RPC communication
  • Silent mode (no console output interfering with protocol)
  • Full CodeMie proxy integration for enterprise SSO
  • Profile-based configuration support
  • All Claude Code capabilities via ACP protocol

IDE Configuration:

After installation, configure your IDE to use codemie-claude-acp:

Zed (~/.config/zed/settings.json):

{
  "agent_servers": {
    "claude": {
      "command": "codemie-claude-acp",
      "args": ["--profile", "work"]
    }
  }
}

JetBrains (~/.jetbrains/acp.json):

{
  "agent_servers": {
    "Claude Code via CodeMie": {
      "command": "codemie-claude-acp",
      "args": ["--profile", "work"]
    }
  }
}

Emacs (with acp.el):

(setq acp-claude-command "codemie-claude-acp")
(setq acp-claude-args '("--profile" "work"))

Usage:

# Install the ACP adapter
codemie install claude-acp

# The command is typically invoked by your IDE, not directly
# But you can test it manually:
codemie-claude-acp --profile work

Note: Unlike codemie-claude, the ACP adapter is designed to be invoked by editors, not used interactively in a terminal. All output goes through the JSON-RPC protocol.

Gemini CLI

Google's Gemini AI coding assistant with advanced code understanding.

Installation: codemie install gemini

Requirements:

  • Requires a valid Google Gemini API key from https://aistudio.google.com/apikey
  • Requires Gemini-compatible models only (gemini-2.5-flash, gemini-2.5-pro, etc.)
  • LiteLLM or AI-Run SSO API keys will not work with Gemini CLI

Setup:

# Configure Gemini with dedicated API key
codemie setup
# Select: "Google Gemini (Direct API Access)"
# Enter your Gemini API key from https://aistudio.google.com/apikey

# Or use environment variable
export GEMINI_API_KEY="your-gemini-api-key-here"

Features:

  • Advanced code generation and analysis
  • Multi-model support (Gemini 2.5 Flash, Pro, etc.)
  • Project-aware context with directory inclusion
  • JSON and streaming JSON output formats

Usage:

codemie-gemini                          # Interactive mode
codemie-gemini "your prompt"            # With initial message
codemie-gemini -p "your prompt"         # Non-interactive mode (Gemini-specific)
codemie-gemini -m gemini-2.5-flash      # Specify model (Gemini-specific)
codemie-gemini --model gemini-2.5-flash "analyze code"  # With config override

Note: Installed via Python (pip/uv), not npm. Requires Python 3.9+ and Anthropic or OpenAI API key.

OpenCode

Open-source AI coding assistant with comprehensive session analytics.

Installation: codemie install opencode

Requirements:

  • Node.js 20.0.0 or higher
  • Supported providers: LiteLLM, AI/Run SSO, or direct API access
  • OpenCode CLI installed globally (opencode-ai npm package)

Features:

  • Modern open-source AI assistant
  • Multi-provider support via CodeMie proxy
  • Automatic session metrics extraction
  • Token usage and cost tracking
  • Conversation history preservation
  • SSO authentication support
  • Cross-platform support (Windows, Linux, macOS)

Session Analytics: OpenCode integrates with CodeMie's analytics system to automatically track:

  • Token usage (input, output, total)
  • Session duration and conversation turns
  • Model usage and costs
  • Session metadata and context

All metrics are automatically extracted at session end and viewable via:

codemie analytics --agent opencode

Manual Metrics Processing:

# Process specific session
codemie opencode-metrics --session <session-id>

# Discover and process all recent sessions
codemie opencode-metrics --discover

# Detailed processing output
codemie opencode-metrics --discover --verbose

Session Storage: OpenCode sessions are stored following XDG Base Directory Specification:

  • Linux: ~/.local/share/opencode/storage/
  • macOS: ~/Library/Application Support/opencode/storage/
  • Windows: %LOCALAPPDATA%\opencode\storage\

Configuration: CodeMie automatically configures OpenCode to use the selected provider via:

  • Environment variable injection (OPENCODE_CONFIG_CONTENT)
  • Dynamic model configuration in OpenCode's native format
  • SSO proxy routing for enterprise authentication

Example Workflow:

# Install OpenCode
codemie install opencode

# Configure provider (if not already done)
codemie setup

# View metrics
codemie analytics --agent opencode