👋 Introduction

mcp-opencode is an MCP server that lets your AI assistant query any model your local opencode install is configured for — anthropic, github-copilot, google-vertex, and any other provider — without ever handling an API key.

It acts as a thin bridge: your assistant calls a tool, the server forwards the prompt to opencode's local HTTP server, and the response comes back. opencode owns the credentials; this server never sees them.

🔐 The zero-key delegation model

Most MCP servers that talk to model providers need their own API tokens, stored in config or environment variables. mcp-opencode takes a different approach:

• No tokens here — you configure your providers once in opencode (via opencode auth), and this server inherits them automatically.
• No secrets in flight — no credentials are stored in or passed through the MCP server. Authentication is delegated entirely to opencode's own config.
• One source of truth — the models you can query are exactly the models opencode knows about. Run opencode models to see the same list the server exposes.

This means you get a second-opinion model, a cheaper model for bulk work, or a different provider's flagship — all from inside your assistant, with zero extra auth setup.

✨ Features

• Query any model — send prompts to anthropic, github-copilot, google-vertex, and any other provider you've configured in opencode.
• Discover models — list every model opencode knows about, optionally filtered by provider.
• Zero config auth — no API tokens; authentication is delegated entirely to opencode.
• Auto-start — spins up the opencode server automatically if it isn't already running.
• Allow/block filters — restrict which models are accessible via MCP_OPENCODE_MODEL_ALLOW / MCP_OPENCODE_MODEL_BLOCK.
• Works with any opencode provider — whatever you've configured, you can query.

✅ Prerequisites

• opencode installed, with at least one provider configured.
• Node.js ≥ 20.

📦 Install

npm install -g @kud/mcp-opencode

Then register it with your AI client. For Claude Desktop, edit the config:

{
  "mcpServers": {
    "opencode": {
      "command": "npx",
      "args": ["-y", "@kud/mcp-opencode"]
    }
  }
}

The same command / args pair works for Cursor, Windsurf, VSCode (Copilot), and the Claude Code CLI — only the surrounding config file differs.

💬 Example conversations

"What models do I have access to?" → Calls list_models, returns all configured models grouped by provider.

"List only my Anthropic models." → Calls list_models with provider: "anthropic".

"Ask GPT-4.1 via Copilot to explain what a monad is." → Calls query with model: "github-copilot/gpt-4.1".

"Use Claude Sonnet to review this diff and suggest improvements." → Calls query with model: "anthropic/claude-sonnet-4-6" and the diff as the prompt.

"Get a second opinion from Gemini on this architecture decision." → Calls query with model: "google-vertex/gemini-2.5-pro".

Head to Tools for the full reference, or How it works to understand the bridge.