MCP Vector Proxy

You will run a semantic MCP proxy that sits between AI agents and MCP Router, exposing only four tools to simplify tool discovery and execution. It uses local vector embeddings to identify the right tool on demand and does not require an OpenAI key.

How to use

Connect your AI client to the proxy using either the HTTP endpoint or the local stdio bridge. The proxy offers four tools for interaction: discover_tools, execute_tool, batch_execute, and refresh_tools. Your AI agent searches semantically for what it needs and then calls the appropriate tool, reducing token usage while maintaining access to the full MCP ecosystem.

How to install

Prerequisites you need before installation: Node.js 18+ and an active MCP Router instance running on your network.

# 1) Copy the example environment setup
cp .env.example .env
# Edit .env and replace the placeholder with your real MCP Router token

# 2) Install dependencies and build the project
npm install
npm run build

# 3) Register auto-start and launch the tray on your platform
# Windows
npm run setup
# macOS / Linux
npm run setup

# 4) Connect your AI clients by configuring their MCP server reference
# Example for Claude Code

Additional content

Configuration notes and how the proxy communicates with MCP Router are provided below. The proxy exposes both an HTTP endpoint for remote clients and a local stdio bridge for desktop agents. It indexes all MCP tools, uses embedding-based semantic search combined with keyword matching, and auto-syncs when tools change.

Available tools

discover_tools

Perform a hybrid semantic and keyword search to find relevant MCP tools based on a natural language query.

execute_tool

Run a specific MCP tool by its exact name with provided arguments.

batch_execute

Execute multiple MCP tools in parallel within a single call.

refresh_tools

Force a full re-index of all tools from the MCP Router to refresh embeddings and metadata.