👻 GhostCoder

The AI coding partner that doesn't trust itself.

Why GhostCoder?

	Cursor	Copilot	GhostCoder
UI	Sidebar chat	Inline completion	No UI. Inline hints only
Trust	You trust the AI	You trust the AI	AI challenges itself
Audit	None	None	Full replay + explain
Safety	None	None	Guardrails block dangerous
Cost	$20/mo	$10/mo	Free, local, MIT
Hardware	Cloud	Cloud	GTX 1650 4GB

Three Features Nobody Else Has

🔍 Inline Hints

Specialist agents watch your code and whisper fixes. No sidebar. No context switching.

When you run a command that fails, or write code with an obvious bug, GhostCoder classifies the error and routes it to the correct specialist agent (e.g., security engineer, database optimizer, frontend developer) to give you a dismissible inline hint.

🔁 Ghost Replay

Every decision is recorded, explainable, and replayable.

If you want to understand why GhostCoder suggested a particular change or see a step-by-step trace of how an error was handled, you can inspect it or run a full replay.

# Explain what happened in a specific session event
ghostcoder explain --session 20260716-001 --event 0

# Generate a project automation report (weekly/monthly)
ghostcoder report --period week

# Re-apply proposed fixes from a session
ghostcoder replay --session 20260716-001

🛡️ Ghost Skeptic & Safety Guardrails

GhostCoder is the first coding assistant that validates its own output.

• Ghost Skeptic: Uses an adversarial classifier model to look for logic flaws and security issues (e.g. plaintext credentials) in every suggested fix. If a flaw is detected, it is either corrected or blocked.
• Safety Guardrail: Automatically screens proposed code blocks for database-destructive operations (DROP DATABASE) or risky directory deletions (rm -rf) to keep your workspace safe.

---

Model Agnostic & GPU Auto-Scaling

GhostCoder is completely model-agnostic and works with any Large Language Model supported by your local Ollama instance (including Llama 3, DeepSeek, Mistral, Qwen, Codellama, etc.).

By default, GhostCoder automatically detects your local GPU resources and dynamically routes requests to the largest model that fits comfortably in your VRAM budget (up to 85% utilization).

GPU Tiers & Recommendations

VRAM Capacity	Tier Profile	Classifier Model	Coder Model	Reasoner Model	Skeptic Model
< 6 GB	entry	qwen2.5:0.5b	qwen2.5-coder:1.5b	qwen2.5-coder:1.5b	qwen2.5:0.5b
6 - 12 GB	mid	qwen2.5:0.5b	qwen2.5-coder:7b	qwen2.5-coder:7b	qwen2.5:0.5b
12 - 20 GB	high	qwen2.5:0.5b	qwen2.5-coder:14b	deepseek-coder:6.7b	qwen2.5:3b
20 - 48 GB	workstation	qwen2.5:3b	qwen2.5-coder:32b	deepseek-coder:33b	qwen2.5:7b
>= 48 GB	datacenter	qwen2.5:7b	qwen2.5-coder:72b	deepseek-coder:33b	qwen2.5:14b

Customizing Your Model Configuration

You can hot-swap any role to run any custom model pulled locally:

# Configure coder model to use Llama 3
ghostcoder config --model-coder llama3:8b

# Configure skeptic model to use Mistral
ghostcoder config --model-skeptic mistral:7b

---

Quick Start

1. Prerequisites

Install Ollama and pull the models you want to use. For the default Entry configuration:

ollama pull qwen2.5:0.5b
ollama pull qwen2.5-coder:1.5b

2. Install GhostCoder

Install the CLI and background daemon:

pip install ghostcoder

3. Initialize Integrations

Configure your shell and start the daemon:

ghostcoder init

This automatically appends source lines to your shell configuration (.bashrc, .zshrc, or config.fish) and launches the background daemon.

---

Editor & IDE Setup

Neovim

Add the plugin folder to your runtimepath (e.g., using your package manager):

-- lazy.nvim configuration example
{
  "anujsingh-cse/GhostCoder",
  config = function()
    -- Integration works automatically over the Unix Domain Socket
  end
}

VS Code

1. Open the vscode folder in this repository.
2. Package the extension:

   cd vscode
   npm install
   npm run compile
   npx vsce package

3. Install the generated .vsix file in VS Code.

---

CLI Command Reference

• ghostcoder start - Start the background daemon.
• ghostcoder status - Show active models, VRAM usage, and active integrations.
• ghostcoder logs - Tail background daemon log output.
• ghostcoder stop - Stop the daemon gracefully.
• ghostcoder skeptic --on|--off - Enable or disable the skeptic validation agent.

---

License

This project is licensed under the MIT License - see the LICENSE file for details.