Elixium Forge ⚒️

Your AI. Your Hardware. Your Code Stays Home.

Elixium Forge is an autonomous AI coding agent that lives inside VS Code. It reads your codebase, plans changes, writes files, and executes commands — powered entirely by local models on your own hardware. No cloud. No API keys. No data leaving your network.

Part of the Elixium ecosystem.

The Elixium Ecosystem

Elixium              Cloud AI platform           elixium.ai
Elixium Companion    VS Code → Cloud Elixium     marketplace
Elixium Forge        VS Code → Local models      ← you are here

Quick Start

You need:

A model server — Ollama, LM Studio, or any OpenAI-compatible endpoint
A model — qwen2.5-coder:14b, gemma-4, deepseek-coder, or your pick

Setup:

Install Elixium Forge from the VS Code Marketplace
Settings → search "Elixium"
Set Server URL → http://localhost:11434 (or your LAN server with /v1 for OpenAI-compatible)
Set Default Model → your model name
Open the Elixium Forge sidebar → start building

Auto-detects Ollama native vs OpenAI-compatible from the URL. Works with any model that speaks either protocol.

How It Works

Three-Pass Agent System

Most AI coding tools ask the model to think AND write structured tool calls in the same breath. Local models struggle with this. Forge separates the concerns:

Pass 1 — Think     ReAct loop reads files, greps code, plans the change
                    The model does what it's good at: reasoning
                    
Pass 2 — Apply     Separate prompt: "Here is the file. Here are the changes.
                    Output ONLY the new file." No XML tags needed.
                    
Pass 3 — Audit     Read the file back. Verify it exists and has content.
                    Report the result.

Inspired by Cursor's two-model architecture, adapted for single-model local inference.

11 Tools

	Tool	What It Does
🔍	GREP	Search file contents (ripgrep with Node.js fallback)
📂	GLOB	Find files by pattern
📖	READ_FILE	Read with line numbers, optional offset/limit for large files
📁	LIST_DIR	Workspace file tree
✏️	EDIT	Safe edit — requires read-first, enforces uniqueness
📝	CREATE_FILE	Write new files with path traversal protection
🔄	SEARCH_REPLACE	Legacy find-and-replace
🗑️	DELETE_FILE	Moves to trash, permission required
▶️	RUN_COMMAND	Shell execution with permission system
🔧	INIT_REPO	Git init + optional GitHub publish
🎨	GENERATE_IMAGE	Fooocus API integration (optional)

Agent Modes

Mode	Tools	Use Case
Agent	All 11	Autonomous multi-step tasks
Plan	Read-only	Analysis without modifications
Ask	None	Conversational Q&A

What Makes Forge Different

vs. GitHub Copilot / Cursor — Forge runs 100% offline on your hardware. Your code never leaves your network.

vs. Cline / Roo Code — Forge doesn't demand perfect XML tool calls from local models. The fuzzy parser catches whatever format the model throws (call:TOOL{...}, JSON blocks, Python-style calls), and the three-pass system means the model never needs to produce structured output for file writes.

vs. ChatGPT / Claude in browser — Forge has direct workspace access. It reads your actual files, runs your actual commands, writes to your actual disk. It's not a chat window — it's an agent.

Safety

Permission system — writes, commands, and destructive actions require approval
Dangerous command blocking — rm, sudo, chmod blocked by default
Read-first enforcement — must READ_FILE before EDIT
Edit uniqueness check — refuses edits that match multiple locations
Path traversal protection — can't write outside workspace
Checkpoint system — asks "keep going?" every 50 turns

Configuration

Settings → search "Elixium":

Setting	What	Example
Server URL	Model server endpoint	`http://localhost:11434`
API Key	Optional, for authenticated servers	leave blank for Ollama
Default Model	Main coding model	`qwen2.5-coder:14b`
Fast Model	Chat, quick tasks	`qwen2.5:7b`
Deep Model	Planning, complex reasoning	`qwen2.5:32b`

"elixium.commandExecutionMode": "interactive",
"elixium.agentMode": "react",
"elixium.maxAgentTurns": 100,
"elixium.allowDangerousCommands": false

Troubleshooting

Problem	Fix
Server unreachable	Check Server URL in settings. Use IP for LAN servers.
Model not responding	Verify model is loaded: `curl http://server:11434/api/tags`
Agent claims it edited but didn't	Three-pass system should handle this. Check Output panel.
Extension not updating	Uninstall → Cmd+Q → reinstall VSIX
Chat stuck	Click 🗑️ in sidebar header

License

MIT · Built by IndirectTek for the Elixium ecosystem.