CodeForge

An open-source, model-independent, memory-native, agentic coding IDE.

CodeForge is an AI-native code editor built on Code-OSS. One Windows install, your own models (cloud or local), no Docker required.

Also discoverable via: Cursor AI, Cursor-style AI IDE, open source AI coding IDE, Ollama coding agent.

Download: Latest Windows installer · Source of truth: extensions/codeforge/

Why CodeForge?

	CodeForge
Models	OpenAI, Anthropic, Gemini, xAI, OpenRouter, Ollama, vLLM, LM Studio, any OpenAI-compatible API
Agent	Multi-file agent with approvals; composer modes Ask / Plan / Agent + permission chip (Default → Assisted → Allow all)
Chat UI	Cursor-like sidebar: footer composer, context ring, clean tool rows, Undo/Review, Restore checkpoints
Weak local models	Opt-in small-model harness (codeforge.ai.weakModelMode: on/off)
Sandbox	Configurable shell sandbox + optional Full agent freedom
Memory	Session wiki, project wiki, stable facts on disk (~/.CodeForge/)
Governance	Editable skills, rules, policies, hard guardrails
MCP	Inbound tools + outbound wiki server + Plugins catalog

Install (Windows)

1. Open Releases
2. Download CodeForge-Setup-*-win32-x64.exe
3. Run the installer (per-user under %LOCALAPPDATA%\Programs\CodeForge)

Or build from source (below).

Quick Start (development)

git clone https://github.com/Kortexio/CodeForge.git
cd CodeForge
npm install
npm run vscode:setup      # clone Code-OSS, brand, install deps
npm run build:extensions # compile AI extension
npm run vscode:compile   # compile editor (first time is slow)
npm run vscode:dev       # launch CodeForge with AI sidebar

See docs/code-oss-integration.md for details.

Features

• Embedded AI Platform — agent, chat, settings, and tools inside the IDE
• Cursor-like composer — Ask / Plan / Agent in the chat footer; Shift+Tab cycles mode; permissions chip separate from mode (Allow all ≈ former Auto)
• Context ring — live prompt tokens vs window limit on the composer
• Tool timeline — short Cursor-style labels (Shell npm test, Read path), exploration grouping, failures stay expanded
• Edit journal — Undo All / Review Changes (vscode.diff against pre-edit snapshots)
• Restore checkpoint — revert a user turn (messages + related tool traces + journaled files)
• Shell sandbox — codeforge.ai.shellSandbox (restricted → unrestricted) and Full agent freedom in Agent settings
• Model Independence — bring your own API keys or local servers
• Plan mode — explore + write a wiki plan without mutating source
• Small-model harness — opt-in phase tools + tighter guardrails for small/local LLMs
• Persistent Memory — session/project wiki and temporal facts
• Code Intelligence — hybrid search (lexical + semantic embeddings)
• MCP + Plugins — inbound tools, outbound wiki server, marketplace-style plugins catalog
• Tab Completion — inline completions with your model
• Browser Agent — visual checks (Playwright optional)
• Git Intelligence — status, diff, blame, conflicts, commit messages

Architecture

CodeForge/
├── extensions/codeforge/   # Shipped AI platform (source of truth)
│   ├── src/agent/               # Agent loop, edit journal, tool registry, harness
│   ├── src/views/               # Chat host + chatViewHtml (Cursor-like UI)
│   ├── src/memory/              # Session + project wiki + temporal facts
│   ├── src/storage/             # ~/.CodeForge paths, sessions, artifacts, traces
│   ├── src/context/             # Context budget / ranking
│   ├── src/sandbox/             # Shell isolation + unrestricted / freedom
│   ├── src/intelligence/        # Index, LSP, git, embeddings
│   ├── src/browser/             # Browser agent
│   ├── src/mcp/                 # Inbound + outbound MCP
│   └── src/governance/          # Skills, rules, guardrails, HITL
├── docs/                        # Documentation
├── tests/                       # Test suites
└── scripts/                     # Code-OSS setup / package

See docs/architecture.md for the full component model.

Configuration

Prefer CodeForge Settings (Models / MCP / Guardrails / Agent) in the app.

Legacy mirrors (active server only):

{
  "codeforge.ai.mode": "agent",
  "codeforge.ai.weakModelMode": "off",
  "codeforge.ai.shellSandbox": "restricted",
  "codeforge.ai.fullAgentFreedom": false
}

Supported providers: OpenAI, Anthropic, Google (Gemini), xAI, OpenRouter, Ollama, vLLM, LM Studio, custom OpenAI-compatible endpoints.

Development

Prerequisites

• Node.js matching vscode/.nvmrc (for full IDE builds)
• npm, Git
• On Windows: Visual Studio C++ workload + Spectre-mitigated libs

Building & tests

npm install
npm run build:extensions
npm test

Packaging (Windows)

# Installer + portable + zip (recommended for distribution)
npm run vscode:package:win

# Installer only (from existing portable build)
npm run installer:win

Requires Inno Setup: winget install JRSoftware.InnoSetup

Artifact: dist/codeforge-win32-x64/CodeForge-Setup-<version>-win32-x64.exe

Discoverability

Search on GitHub for CodeForge, AI coding IDE, Cursor AI, or topics below.

• Repo: github.com/Kortexio/CodeForge
• Releases: github.com/Kortexio/CodeForge/releases
• Topics: ide · ai · vscode · code-oss · openai · ollama · agent · mcp · electron · typescript · cursor · cursor-ai · ai-ide · ai-coding · coding-agent · llm

License

MIT License — see LICENSE for details.

Contributing

Contributions are welcome! Please read our Contributing Guide before submitting a Pull Request.

Acknowledgments

• Built on Code-OSS, the open-source foundation of Visual Studio Code
• Inspired by modern AI coding assistants
• Powered by open AI research and models