OpenWire

Expose VS Code language models as an OpenAI-compatible REST API on localhost.

One extension. Every model VS Code can see. Standard API. Built for agents.

---

Features

• OpenAI-compatible — /v1/chat/completions, /v1/models with streaming (SSE)
• Auto-discovery — finds every language model registered in VS Code
• Tool forwarding — pass OpenAI-format tools, get tool_calls back
• Multi-provider content handling — normalises Anthropic-style content arrays, OpenAI strings, and Gemini parts into a consistent format
• XML tool call fallback — when native tool forwarding isn't available, parses Claude's XML <function_calls> output into proper tool_calls objects
• Rate limiting — configurable per-minute request cap
• API key auth — optional Bearer token authentication
• Zero dependencies — pure Node.js HTTP, no Express, no frameworks

Models

Any model available through VS Code's Language Model API is automatically exposed — no configuration needed. This typically includes:

• Claude — Opus, Sonnet, Haiku
• GPT — Codex, GPT-4.1, o4-mini
• Gemini — Gemini Pro, Gemini Flash
• Ollama — any locally running Ollama models (Llama, Qwen, DeepSeek, Mistral, etc.)
• Any other models registered via the VS Code Language Model API

Run GET /v1/models to see what's available in your setup.

Provider Compatibility

OpenWire normalises differences between providers so callers always get a consistent OpenAI-format response:

Provider	Content format	Tool calling	Status
Claude (Anthropic)	Array of {"type":"text","text":"..."} parts	Native via VS Code API; XML <function_calls> fallback parsed automatically	✅ Full support
GPT (OpenAI)	Plain string	Native tool_calls via VS Code API	✅ Full support
Gemini (Google)	Plain string or parts array	Native via VS Code API	✅ Full support
Ollama (local)	Plain string	Depends on model capability	✅ Supported

Content normalisation — Incoming messages with content as an array of content parts (Anthropic format), a plain string (OpenAI/Gemini), or null are all normalised to plain strings before forwarding to the VS Code LM API.

Tool call fallback — When the VS Code LM API can't forward tools natively (e.g. older VS Code versions), Claude may output tool calls as XML. OpenWire detects and converts these to standard tool_calls objects in the response, so callers never see raw XML.

Quick Start

Install from the VS Code Marketplace (or load the .vsix). The server starts automatically on http://127.0.0.1:3030.

# List available models
curl http://localhost:3030/v1/models

# Chat completion
curl http://localhost:3030/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4.6",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

# Streaming
curl http://localhost:3030/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4.6",
    "messages": [{"role": "user", "content": "Explain zero-knowledge proofs"}],
    "stream": true
  }'

Endpoints

Method	Path	Description
GET	/health	Health check
GET	/v1/models	List available models
GET	/v1/models/:id	Get specific model
POST	/v1/chat/completions	Chat completion (streaming + non-streaming)
POST	/v1/completions	Legacy completions (mapped to chat)

Configuration

All settings live under openWire.server.* in VS Code:

Setting	Default	Description
autoStart	true	Start server when VS Code launches
host	127.0.0.1	Bind address
port	3030	Port number
apiKey	""	Bearer token for authentication
defaultModel	""	Fallback model when none specified
defaultSystemPrompt	""	Injected system prompt if none present
maxConcurrentRequests	4	Concurrent request limit
rateLimitPerMinute	60	Rate limit
requestTimeoutSeconds	300	Request timeout
enableLogging	false	Verbose logging

Commands

• OpenWire: Start Server
• OpenWire: Stop Server
• OpenWire: Restart Server
• OpenWire: Toggle Server

Using with OpenClaw

OpenWire can serve as a model provider for OpenClaw agents. Register OpenWire as a custom provider called copilot-proxy in your ~/.openclaw/openclaw.json:

{
  "models": {
    "providers": {
      "copilot-proxy": {
        "baseUrl": "http://localhost:3030/v1",
        "apiKey": "n/a",
        "api": "openai-completions",
        "authHeader": false,
        "models": [
          {
            "id": "claude-sonnet-4.6",
            "name": "Claude Sonnet 4.6",
            "contextWindow": 128000,
            "maxTokens": 8192
          }
          // add any other models from /v1/models
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "copilot-proxy/claude-sonnet-4.6"
      }
    }
  },
  "plugins": {
    "entries": {
      "copilot-proxy": { "enabled": true }
    }
  }
}

Set authHeader: false since OpenWire handles authentication through VS Code's Copilot session — no API keys are needed. Run curl http://localhost:3030/v1/models to see all available model IDs.

Architecture

src/
  extension.ts          — activation, commands, status bar
  models/
    discovery.ts        — model discovery, caching, dedup
  routes/
    chat.ts             — chat completions + tool forwarding
  server/
    config.ts           — settings loader
    gateway.ts          — HTTP server, routing, middleware
  ui/
    sidebar.ts          — webview sidebar panel
  types/
    vscode-lm.d.ts      — type augmentations

Lightweight · zero runtime dependencies

License

MIT