MCP Inline Complete

AI-powered inline code completions for VS Code via a Model Context Protocol server.

Architecture

VS Code Extension (MCP Client)
  └── InlineCompletionItemProvider
        │  debounce + context extraction
        │
        └──── stdio ────► MCP Server
                            └── get_inline_completion tool
                                  └── Anthropic Claude API

Setup

Prerequisites

• Node.js ≥ 18
• ANTHROPIC_API_KEY environment variable set

Install & Build

# Install all dependencies
npm install

# Build everything
npm run build

Development

# Terminal 1: Watch server
cd packages/server && npm run dev

# Terminal 2: Watch extension
cd packages/extension && npm run dev

Run in VS Code

1. Open this repo in VS Code
2. Press F5 → launches Extension Development Host
3. The MCP server starts automatically as a child process
4. Type code → ghost text completions appear after the debounce delay

Configuration

Setting	Default	Description
mcpInlineComplete.enabled	true	Enable/disable completions
mcpInlineComplete.debounceMs	300	Delay before requesting completion (ms)
mcpInlineComplete.prefixLines	100	Lines before cursor sent as context
mcpInlineComplete.suffixLines	30	Lines after cursor sent as context

Toggle

• Command palette: MCP Inline Complete: Toggle On/Off
• Status bar: click the ✦ MCP indicator

Project Structure

packages/
  shared/     ← Shared TypeScript types
  server/     ← MCP server (stdio transport, Anthropic SDK)
  extension/  ← VS Code extension (InlineCompletionItemProvider)

Customization

Swap the LLM

Edit packages/server/src/index.ts — replace the Anthropic call with any other provider (OpenAI, Ollama, local model, etc). The MCP interface stays the same.

Add more tools

The MCP server can expose additional tools alongside get_inline_completion:

• get_hover_info — context-aware hover documentation
• refactor_selection — AI-powered refactoring suggestions
• explain_code — inline code explanations