GitHub - vcaesar/codg: The next easy code and work AI agents harness system: auto, asynchronous, concurrency, efficiently and High accuracy

6 min read Original article ↗

Latest Release Build Status GoDoc Join the Discord chat at https://discord.gg/Dy5QZRbaND

The next easy code and work AI agents harness system, auto and asynchronous, concurrency and high performance, Efficiently and High accuracy.

English | 繁體中文 | 简体中文 | 日本語 | 한국어 | Français | Deutsch | Español | Português | Русский | العربية

Codg Demo Codg Demo

Installation

Mac and Linux:

# Homebrew
brew install vcaesar/tap/codg

# NPM
# npm install -g @vcaesar/codg

Windows (PowerShell):

# Winget
# winget install vcaesar.codg

# YOLO (native PowerShell installer)
irm https://raw.githubusercontent.com/vcaesar/codg/main/demo/boot.ps1 | iex

All (macOS, Linux, or Windows via Git Bash / MSYS2 / Cygwin / WSL):

# YOLO
curl -fsSL https://raw.githubusercontent.com/vcaesar/codg/main/demo/boot.sh | bash

Or click Releases directly to download and run it.

Go to your project directory, run codg, use "/init" to init the projects. Use "/yolo" to toggle the auto and ask mode, and you can set permissions by codg.toml.

Features

  • Auto and asynchronous, concurrency and high performance agents system, and low memory use
  • Multi models providers (40+ API and Pro providers, Custom URL API) and local models via by openai-compat or claude-compat, Support Openrouter, Ollama, Nvidia and others free models, use it by "/connect" "/models" or "codg auth"
  • Asynchronous and model rules for the multi agents
  • Compression input-output, context and prompts to save the tokens, cache and rules to reduce the cost
  • Any terminal and OS support, also web terminal support
  • Easy use: The TUI in everywhere like GUI and Easy, Desktop and Web in the BETA
  • Click or "/xxx" to switch sessions, Click to everywhere in TUI
  • Clcik "Modified Files" or "/diff" and "/diff git" to view the diff files in TUI same the vscode
  • Autocomplete the English letters and short sentences
  • More easy Agents, Skills and MCP system, custom Agents and Skills support
  • Channel and features support like OpenClaw

Desktop App (BETA), Web (BETA), Claw (BETA), Some features need wait for the test and fix bugs then release it.

Providers

Atom, Copilot, Anthropic, Anthropic API, OpenAI, OpenAI API, Gemini, Gemini API, OpenRouter, Antigravity, Cursor, Kiro, xAI, Azure, Bedrock, Vertex AI, Nvidia, HuggingFace, Vercel, Ollama Cloud, Cloudflare Workers, GitHub, Poe, Meta, Groq, IO.net, OpenCode Zen, OpenCode Go, Windsurf, Cerebras.

China: Z.ai, Zhipu, Zhipu Coding, Kimi, Kimi Coding, DeepSeek, MiniMax, MiniMax China, Qwen, MiMo, Qiniu Cloud, Ali Coding, Ali Coding CN, Tencent Coding.

Benchmark

RAM usage

Tool 1 active session 10 active sessions Extra PSS per added session
Codg 65 MB 165 MB ~10 MB
Codex CLI 140.0 MB 334.8 MB ~21.6 MB
Cursor Agent 214.9 MB 1632.4 MB ~157.5 MB
GitHub Copilot CLI 333.3 MB 1756.5 MB ~158.1 MB
OpenCode 371.5 MB 3237.2 MB ~318.4 MB
Claude Code 386.6 MB 2300.6 MB ~212.7 MB

Terminal-bench

SWE

Reporting Bugs:

Open a Github Issues

How we use your data:

Currently no any data and telemetry is collected here, and 100% local model supported, use the API you can see they providers' policies.

For TUI usage, see the TUI commands documentation, and type /help inside the TUI to view key bindings and other help.

CLI Commands

Use: codg -h

codg auth/login               # Authenticate (Atom, OpenAI, GitHub...)
codg web                      # Start web UI on port 4096
codg desktop                  # Launch the desktop app (Wails)
codg claw                     # Start messaging agent (Telegram/Discord/Slack)
codg gateway --private-only   # Start secured gateway
codg models claude            # List models matching "claude"
codg runm start Qwen/Qwen3-8B-GGUF   # Start a local model
codg runm download user/model # Download a GGUF model
codg plugin install repo/name # Install a plugin
codg plugin list              # List installed plugins
codg install repo/name        # Shorthand for plugin install
codg mcp add myserver cmd     # Add an MCP server
codg mcp list                 # List configured MCP servers
codg skill url add <url>      # Add a skill source URL
codg themes set catppuccin    # Switch theme
# codg logs -f                # Tail application logs
codg toml                     # show the all config
codg stats/s                  # Show usage statistics
codg dirs                     # Print data/config directory paths
codg projects                 # List tracked project directories
codg lite 2                   # Set lite mode level (0-4)
codg merge origin main        # Safe git merge with v1/ backup
codg migrate                  # Migrate config from .claude/.opencode
codg vm build                 # Build on remote VM
codg vm run -- make test      # Execute command on VM
codg sandbox run -- ./test.sh # Run in sandbox
codg sandbox status           # Check sandbox availability
codg used                          # Show usage limits and API stats for all providers
codg update                   # Update codg version
codg updatep                  # Update provider definitions

Usage Examples

Non-Interactive (codg run)

# Pipe input from another command.
cat errors.log | codg run "What's causing these errors?"
# Verbose mode (debug output to stderr).
codg run -v "Debug this function"

Web UI

# Start the web UI on default port 4096; (Wait done for the test, then release it).
codg web
# Custom port.
codg web -p 8080

# API-only mode (no frontend, no browser).
codg web 0

Plugin Management

# Install a plugin from a Git repository.
codg install github.com/user/codg-xxx-auth

Custom Agents and Skills:

Copy xx_agent.md (.codg/agents/templates) or SKILL.md (.codg/skills) to the directory

Configuration System

Create a codg.toml in your project root (or ~/.codg/config/codg.toml for global settings):

# codg.toml — Minimal project config.
[options]
lite_mode = 2          # 0 = all agents, 2 = default lean set, 4 = single agent
locale    = "en"       # UI language: en, zh-CN, ja
ctx_resize = true
token_save = 2

[options.tui]
theme     = "catppuccin"
dark_mode = true
compact_mode = false

[tools.grep]
backends = ["rg", "sg", "csearch", "ngram", "regex"]

Provider Setup

# Use an API key (supports $ENV_VAR expansion).
[providers.anthropic]
api_key = "$ANTHROPIC_API_KEY"

# Use OAuth (set via `codg auth`).
[providers.openai]
oauth = true

# Custom / self-hosted provider.
[providers.local]
name     = "My Local LLM"
type     = "openai-compat"
base_url = "http://localhost:8080/v1"
api_key  = "not-needed"

Agent Customization

# Shorthand: assign a model type.
agents.coder = "large"
agents.task  = "small"

# Full form: fine-tune an agent.
[agents.advisor]
model           = "large"
temperature     = 0.3
thinking_budget = 32000

MCP Servers

# HTTP MCP server.
[mcp.websearch]
type = "http"
url  = "https://mcp.exa.ai/mcp?tools=web_search_exa"

Skills

# Auto load and download in TUI or codg skill
[option]
skill_urls = ["https://github.com/user/skills"]

Local Models (llama.cpp)

[llama]
port     = 8090
host     = "127.0.0.1"
ctx_size = 32000
gpu      = "auto"          # auto, cuda, off

Messaging Channels

[channels.telegram]
enabled     = true
token       = "$TELEGRAM_BOT_TOKEN"
allowed_ids = ["123456789"]

[channels.discord]
enabled  = true
token    = "$DISCORD_BOT_TOKEN"

Permissions

[permissions]
allowed_tools = ["bash", "edit", "view", "glob", "grep"]
allowed_dirs = ["**x"] # all directories