Ghost

AI coding assistant for VS Code.

Ghost runs chat, agent tools, and inline completion against Ollama, MLX/VLM, or another OpenAI-compatible server. Your code stays on your machine when you use a local provider.

Highlights

• Local chat through the Ghost view and the native @local chat participant.
• Ollama support with automatic API compatibility handling.
• MLX/VLM and generic OpenAI-compatible provider support.
• Inline code completion with a fast Fill-in-the-Middle model.
• Workspace context from the workspace, folders, active editor, selection, and open files.
• Text-file attachments for focused requests.
• Five workspace tools with allowlists, denylists, approval prompts, and visible progress.
• Source-editor edit previews with Accept/Reject controls, selected-hunk approval, and restore actions for applied edits.
• Ask, Edit, Agent, Explain, and Inline workflow modes.
• Conversation history popup, prompt history, reusable presets, import/export, retry, regenerate, and cancellation.
• Configurable assistant name, avatar, accent color, layout, context, and response behavior.
• Keyboard-friendly interface with visible focus, screen-reader status updates, high-contrast theme support, reduced motion, and narrow-panel layout support.
• No Ghost telemetry service. Provider requests go only to the configured endpoint.

Requirements

• VS Code 1.125 or newer
• Node.js 20 or newer and npm when running Ghost from source
• Ollama, MLX/VLM, or another OpenAI-compatible local server

For the default Ollama setup, install Ollama and pull the recommended models:

ollama pull qwen2.5-coder:7b
ollama pull qwen2.5-coder:1.5b

Start Ollama with its normal desktop app or service. Ghost uses http://localhost:11434 by default.

Install from source

git clone https://github.com/MickyBalladelli/Ghost.git
cd Ghost
npm install
npm run compile

Open the folder in VS Code and press F5 to launch the Extension Development Host. Open Ghost from the Activity Bar or run Ghost: Open Interface from the Command Palette.

Install from a VSIX

Build a VSIX:

npm install
npm run vscode:prepublish
npx vsce package --out ghost.vsix

Then install it from the VS Code Extensions view with Install from VSIX, or run:

code --install-extension ./ghost.vsix

Using Ghost

The Ghost view contains the provider and model strip, message log, composer, and top-bar conversation history.

• Press Enter to send a message. Press Shift+Enter for a new line.
• Use Context to choose workspace, folders, active file, selection, open files, or tools. The popup also shows the available tool and open-file details.
• Use Attach to add text files to the current request. Attachments are not saved in conversation history.
• Use the History button at the top to search, select, rename, or delete previous conversations.
• Use the Settings gear to change the model, workflow, context, appearance, persistence, and tool permissions.
• Choose Agent — implement changes when Ghost should inspect files and modify the workspace. Ghost stages proposed file changes in the real source editor and asks for approval before saving file edits or running terminal commands.
• Use @local in the VS Code Chat view for the native local participant.
• Use @workspace with a keyword to search workspace files and include matching code in the prompt.
• File writes, structured edits, and terminal commands require confirmation unless already approved for the current session.
• Structured edits appear directly in the source editor. Use the Accept Ghost edit or Reject Ghost edit code lens at the top of the file, or use the approval controls in the Ghost view. Accepted edits are saved; rejected edits are restored. Selected hunks can be applied from the Ghost view.

Workspace tools

Ghost exposes these tools to the agent:

The Context popup shows the tools currently available to the request. The Settings panel and ghost.toolAllowlist / ghost.toolDenylist settings control which tools Ghost may use.

Settings

All VS Code settings use the ghost prefix. Open Settings and search for Ghost, or use the Settings gear in the Ghost view.

Provider and model settings

Response settings

Agent permissions

Persistence and diagnostics

The Settings panel also includes:

• Automatic context collection
• Workspace-specific settings
• Context token limit, temperature, response length, and workflow mode
• Provider endpoint and model selection
• Composer height and prompt row count
• Assistant name, avatar, and accent color
• Compact conversation layout
• Provider reasoning visibility
• Thinking details and tool progress visibility
• Telemetry-free diagnostics
• Custom system instructions
• Conversation persistence
• Local debug logging

The panel also manages prompt presets. Saving a preset closes the panel. When conversation persistence is enabled, presets are stored in VS Code extension storage rather than in the project. The Use workspace-specific settings option controls where Ghost configuration is written: enabled writes to the current workspace settings, while disabled writes to global VS Code user settings.

Persistence and privacy

ghost.enableConversationPersistence is off by default. When enabled, Ghost saves conversations and preferences in VS Code storage. State is schema-versioned and common credentials are redacted before storage, export, diagnostics, and display.

Ghost has no telemetry service. Requests are sent only to the provider URL selected in Ghost. If you use a remote OpenAI-compatible or MLX endpoint, the interface marks the connection as external.

Commands

Run these from the Command Palette:

• Ghost: Open Interface
• Ghost: Focus Interface
• Ghost: Check Provider Connection
• Ghost: Check Required Models
• Ghost: Toggle Autocomplete
• Ghost: Reset Interface — delete all Ghost conversations and preferences after confirmation
• Ghost: Export Interface
• Ghost: Clear Interface

There are no default keyboard shortcuts. Add your own in Keyboard Shortcuts if desired. Click the Ghost status bar item to check provider health.

Troubleshooting

Provider is offline

Open the Settings gear, confirm the provider URL, and use Test provider connection. For Ollama, confirm the service is running.

Model is missing

Refresh models in the Ghost view and select an installed model. For the recommended Ollama setup, run:

ollama pull qwen2.5-coder:7b
ollama pull qwen2.5-coder:1.5b

Responses are slow

Lower ghost.maxContextTokens, disable automatic context, hide thinking and tool progress details, or enable compact layout. Long histories are loaded in pages.

A tool is blocked

Check the tool allowlist and denylist in Settings or VS Code settings. File edits and terminal commands still need approval.

OpenAI-compatible endpoint fails

Check the API mode and endpoint suffix. Most OpenAI-compatible servers use /v1; Ollama normally uses its base URL without /v1.

Development

npm install
npm run compile

Create and install a local VSIX with:

./create-vsix.sh

The script increments the patch version, compiles Ghost, creates ghost-${version}.vsix, and installs it in VS Code. Publish an already-created matching VSIX with:

./publish.sh

publish.sh reads the version from package.json, requires ghost-${version}.vsix, and publishes that exact package through vsce.

Run npm test in a VS Code-capable environment for the extension-host test suite. See docs/architecture.md for the extension host, webview, providers, persistence, tools, and approval flow.

Project structure

src/
├── agent/       Chat participant, context injection, and local tool calls
├── providers/   Inline completion provider
├── services/    Ollama, MLX/VLM, and OpenAI-compatible clients
├── tools/       Workspace files, terminal, edits, and tool registration
├── ui/          VS Code webview host, protocol, state, and status bar
└── webview/     Chat interface and interaction logic

License

MIT. See LICENSE.

Links

• Issues and feature requests
• Architecture guide