⚡ MCP Router

Model-agnostic MCP router for the Big Three: OpenAI, Claude, and Gemini.

One API, one UI — point any MCP server at any frontier model. Switch between GPT-5, Claude Opus 4, and Gemini 2.5 Pro without changing your tool integrations.

How It Works

Architecture

┌─────────────┐     ┌──────────────────────┐     ┌───────────────┐
│   Chat UI   │────▶│   POST /api/chat     │────▶│  MCP Servers  │
│  (browser)  │◀────│   Model Router       │◀────│  (any remote) │
└─────────────┘     └──────────────────────┘     └───────────────┘
                           │
                    ┌──────┼──────┐
                    ▼      ▼      ▼
               ┌────────┐ ┌─────┐ ┌────────┐
               │ OpenAI │ │Claude│ │ Gemini │
               │Response│ │ Msg │ │ GenAI  │
               │  API   │ │ API │ │  SDK   │
               └────────┘ └─────┘ └────────┘

Provider Adapters

Provider	API	MCP Strategy	How It Works
OpenAI	Responses API	Server-side	Pass `type: "mcp"` + server URL → OpenAI handles everything
Anthropic	Messages API	Client-side	Connect to MCP, discover tools, agentic loop (tool_use → callTool → tool_result)
Google	GenAI SDK	Client-side	Connect to MCP, discover tools, agentic loop (functionCall → callTool → functionResponse)

Model Auto-Detection

The router detects the provider from the model string:

gpt-*, o3*, o4* → OpenAI
claude-* → Anthropic
gemini-* → Google

Or use explicit prefixes: openai/gpt-4.1, anthropic/claude-sonnet-4-5, google/gemini-2.5-flash

Setup

1. Set API Keys

Add environment variables for the providers you want to use:

Key	Provider	Get it from
`OPENAI_API_KEY`	OpenAI	platform.openai.com
`ANTHROPIC_API_KEY`	Anthropic	console.anthropic.com
`GOOGLE_API_KEY`	Google	aistudio.google.com

You only need keys for providers you plan to use. The UI shows which are configured.

2. Add MCP Servers

In the sidebar, add any remote MCP server URL. Examples:

Val Town MCP: https://www.val.town/mcp
Notion MCP: https://mcp.notion.com/mcp
Any Streamable HTTP or SSE MCP server

3. Chat

Pick a model, type a message. The router handles the rest.

API

POST /api/chat

{
  "model": "gpt-4.1-mini",
  "messages": [
    { "role": "user", "content": "What tools do you have?" }
  ],
  "mcpServers": [
    { "url": "https://mcp.example.com/sse", "label": "my-server" }
  ],
  "system": "You are a helpful assistant."
}

Response (same shape for all providers):

{
  "content": "I have access to the following tools...",
  "toolCalls": [
    {
      "name": "list_tools",
      "serverLabel": "my-server",
      "arguments": "{}",
      "output": "..."
    }
  ],
  "provider": "openai",
  "model": "gpt-4.1-mini",
  "usage": { "inputTokens": 150, "outputTokens": 89 }
}

GET /api/models

Returns available models and which API keys are configured.

Available Models

OpenAI (Responses API — server-side MCP)

gpt-5 — GPT-5 flagship
gpt-4.1 / gpt-4.1-mini / gpt-4.1-nano — GPT-4.1 family
o3 / o4-mini — Reasoning models

Anthropic (Messages API — client-side MCP)

claude-opus-4-6 — Claude Opus 4
claude-sonnet-4-5-20250929 — Claude Sonnet 4.5
claude-haiku-3-5-20241022 — Claude Haiku 3.5

Google (GenAI SDK — client-side MCP)

gemini-2.5-pro — Gemini 2.5 Pro
gemini-2.5-flash — Gemini 2.5 Flash
gemini-2.0-flash — Gemini 2.0 Flash

Key Differences Between Providers

OpenAI is the simplest: their Responses API handles MCP server-side. You pass a URL, they call the MCP server, execute tools, and return results. Zero orchestration code needed.

Anthropic and Google require client-side MCP orchestration. The router connects to MCP servers, discovers tools, converts schemas to each provider's format, then runs an agentic loop — sending messages, catching tool calls, executing them against MCP servers, and feeding results back until the model is done.

Tech Stack

Runtime: Val Town (Deno)
Framework: Hono
OpenAI: openai npm package — Responses API
Anthropic: @anthropic-ai/sdk — Messages API
Google: @google/genai — GenAI SDK
MCP: @modelcontextprotocol/sdk — Streamable HTTP client

Built on Val Town. March 2026.

dcm31

mcp-router