⚒ CodeForge

An agentic coding CLI tool that lets developers interact with their codebase through natural language. It feels like a native CLI tool, not a chatbot wrapper.

Features

• Natural Language Coding — Write, refactor, and debug code by describing what you want
• Multi-Provider AI — Anthropic (default), AWS Bedrock, Google Vertex AI, Azure AI Foundry, NVIDIA, Gemini, OpenRouter
• Plugin System — Extensible via commands, agents, hooks, and skills
• Git Workflows — Automated commits, PRs, branch management
• Security Scanning — Real-time detection of vulnerabilities in code edits
• Session Persistence — Resume any previous session, gzip-compressed
• Context Management — Auto-compacts at 80% of context window
• Permission Controls — Interactive, accept-edits, or bypass modes

Quick Start

# Install
npm install -g codeforge

# Set your API key
export ANTHROPIC_API_KEY=sk-ant-...

# Start coding
codeforge

Installation

npm (recommended)

npm install -g codeforge

Unix (curl)

curl -fsSL https://codeforge.dev/install.sh | bash

Windows (PowerShell)

irm https://codeforge.dev/install.ps1 | iex

Homebrew

brew install codeforge

Usage

Interactive Mode

codeforge

One-shot Mode

codeforge "add error handling to the auth middleware"

Options

-m, --model <tier>        Model: opus, sonnet, haiku (default: sonnet)
-p, --provider <name>     Provider: anthropic, bedrock, vertex, azure, nvidia, gemini, openrouter
--mode <mode>             Permission mode: interactive, accept-edits, bypass
--resume <sessionId>      Resume a saved session
--no-stream               Disable streaming
--verbose                 Verbose logging
--base-url <url>          Custom API base URL

Built-in Commands

| Command | Description | |---------|-------------| | /help | Show all commands | | /develop <feature> | 7-phase feature development workflow | | /review | Run parallel code review | | /commit | AI-generated commit message | | /ship | Commit → push → create PR | | /cleanup | Remove stale branches | | /create-hook | Create custom hook rules | | /iterate | Enable self-improvement loop | | /model <tier> | Switch model tier | | /save | Save current session | | /clear | Clear context |

Agents

Invoke agents with @:

@code-explorer analyze this codebase
@code-architect design a caching layer
@code-reviewer check my latest changes

Plugin System

Plugins live in folders and bundle any combination of:

• Commands — Markdown files with YAML frontmatter (slash commands)
• Agents — Markdown definitions with system prompts
• Hooks — Event handlers (PreToolUse, PostToolUse, SessionStart, SessionEnd, Stop)
• Skills — Reference docs in 3 tiers (description, SKILL.md, examples)

Plugin Locations

1. Built-in: plugins/ directory
2. User: ~/.codeforge/plugins/
3. Project: .codeforge/plugins/

Configuration

Layered precedence (highest wins):

1. Managed/enterprise — ~/.codeforge/managed.json
2. Project local — .codeforge/config.json
3. User home — ~/.codeforge/config.json
4. Environment variables
5. Defaults

Environment Variables

ANTHROPIC_API_KEY        # API key for Anthropic
CODEFORGE_PROVIDER       # AI provider name
CODEFORGE_MODEL          # Model tier
CODEFORGE_BASE_URL       # Custom base URL
CODEFORGE_PERMISSION_MODE # Permission mode
CLAUDE_CONFIG_DIR        # Override config directory

Managed Settings (Enterprise)

{
  "managed": {
    "disableBypass": true,
    "blockedTools": ["WebSearch", "WebFetch"],
    "requireApproval": ["Bash"],
    "trustedHookSources": ["./approved-plugins"]
  }
}

MCP Server Support

Configure MCP servers in your config:

{
  "mcp": {
    "servers": [
      {
        "name": "my-server",
        "transport": "stdio",
        "command": "npx",
        "args": ["-y", "my-mcp-server"]
      }
    ],
    "defaultTimeoutMs": 30000
  }
}

Supported transports: stdio, sse, http, ws

Reliability

• Retry: Up to 6 retries with exponential backoff + jitter
• Rate limits: Respects Retry-After on 429 responses
• Stream timeout: Aborts stalled streams after 5 minutes
• Fallback: Falls back to non-streaming on persistent timeouts
• Context: Auto-compacts at 80% of window via recursive summarization
• Sessions: Persisted as gzipped JSON, fully resumable

Telemetry

Optional OpenTelemetry tracing with sensitive data redaction:

{
  "telemetry": {
    "enabled": true,
    "endpoint": "http://localhost:4318/v1/traces",
    "redactSensitive": true
  }
}

License

MIT