smart-mcp-proxy/mcpproxy-go

Supercharge AI Agents, Safely

What it solves

MCPProxy addresses the limitations of AI agents when connecting to numerous Model Context Protocol (MCP) servers. It solves the problem of hitting tool limits in IDEs (like Cursor's 40-tool limit) and LLM APIs (like OpenAI's 128-function cap), while reducing the massive token cost associated with loading hundreds of tool schemas into an agent's context window.

How it works

It acts as a secure gateway between AI agents and multiple upstream MCP servers. Instead of the agent seeing every available tool, MCPProxy provides a small set of built-in tools. The agent uses a retrieve_tools function to search for needed tools via BM25 search; MCPProxy then returns only the most relevant matches. The agent only loads full schemas for the specific tools it intends to use. Additionally, it implements a security layer that quarantines new servers and allows for scanning via pluggable Docker-based security scanners (e.g., Snyk, Semgrep) before approval.

Who it’s for

Developers and AI power users who use MCP-compatible tools (Cursor, VS Code, Claude Desktop, Goose) and need to scale their agent's capabilities across hundreds of tools without sacrificing performance, token budget, or security.

Highlights

  • Token Efficiency: Reduces token usage by up to 99% by loading only required tool schemas on demand.
  • Security Quarantine: Automatically blocks "Tool Poisoning Attacks" by quarantining new servers until manually approved.
  • Pluggable Scanning: Integrates with Docker-based scanners like Snyk and Trivy to evaluate quarantined servers.
  • Cross-Platform: Distributed as a single binary for macOS, Windows, and Linux with an embedded web UI.
  • Preflight Checks: Allows headless automations to verify tool availability before starting expensive agent sessions.

Related

  • Project
    bvisible/mcp-ssh-manager

    An MCP server that enables AI agents to securely manage remote servers via SSH, featuring per-server security modes to control command execution.

  • Dispatch
    Code Execution with MCP: Building More Efficient AI Agents

    Anthropic introduces a method for AI agents to interact with Model Context Protocol (MCP) servers via code execution rather than direct tool calls, reducing token consumption and improving efficiency.

  • Project
    cloudflare/mcp

    A token-efficient MCP server that enables AI agents to interact with the Cloudflare API using a code-execution pattern to avoid prompt bloat.

  • Project
    brightdata/brightdata-mcp

    Bright Data MCP is a Model Context Protocol server that gives AI agents and LLMs real-time access to public web data through 32 tools for search, scraping, structured extraction from 1,200+ pre-built scrapers, and remote browser automation, all routed through Bright Data's anti-blocking infrastructure.

  • Project
    RekklesNA/ProxmoxMCP-Plus

    ProxmoxMCP-Plus is a bridge that lets AI agents (via MCP) and standard HTTP tools (via OpenAPI) safely operate a Proxmox VE virtualization server, covering VM/LXC lifecycle, snapshots, backups, ISOs, container commands, and persistent job tracking. It emphasizes security with API tokens, command policies, approval gates, and TLS validation, and supports local stdio, Docker-based HTTP, and an opt-in sandboxed Code Mode. An optional OAuth browser-gate flow backed by PostgreSQL handles remote client authentication.