smart-mcp-proxy/mcpproxy-go

What it solves

MCPProxy addresses the limitations of Model Context Protocol (MCP) implementations in AI agents, specifically bypassing tool limits (such as Cursor's 40-tool or OpenAI's 128-function caps) and reducing token consumption. It prevents context window bloat by ensuring agents only load the schemas of tools they actually need rather than hundreds of unused ones, while adding a security layer to protect against tool poisoning attacks.

How it works

It acts as a safe proxy endpoint between the AI agent and multiple upstream MCP servers. Instead of the agent seeing every tool, it uses a retrieve_tools function to perform a BM25 search for relevant tools based on keywords. The agent then loads full schemas only for the selected tools via describe_tool. To execute a tool, the agent uses intent-based variants (call_tool_read, call_tool_write, or call_tool_destructive), which the proxy validates against tool annotations and security policies before execution.

Who it’s for

Developers and power users of AI agents (e.g., Cursor, VS Code, Claude Desktop, Goose) who manage a large number of MCP servers and require advanced security, auditing, and token optimization.

Highlights

• Token Efficiency: Reduces token usage by up to 99% and improves accuracy by 43% by loading tools on demand.
• Security Quarantine: Automatically blocks new servers until manually approved and supports pluggable Docker-based scanners (Snyk, Semgrep, Trivy).
• Cross-Platform: Distributed as a single binary for macOS, Windows, and Linux with an embedded web UI.
• Auditability: Maintains a local activity log of every tool call for full reconstruction of agent workflows.
• Preflight Checks: Allows headless automations to verify tool availability and health before starting expensive agent sessions.