Skip to main content

Introduction

mcpzip is an MCP proxy that aggregates multiple upstream MCP servers and exposes them through a Search + Execute pattern. Instead of loading hundreds of tool schemas into your AI's context window, mcpzip gives the model just 3 meta-tools to discover and invoke any tool on demand.

The Problem​

The Model Context Protocol (MCP) lets AI assistants use external tools. But each MCP server you add dumps all its tool schemas into the context window. With 5+ servers, you're looking at hundreds of tool definitions loaded on every single message:

  • Context bloat: Tool schemas eat up precious context tokens
  • Slower responses: More context = higher latency
  • Model confusion: Too many irrelevant tools degrade tool selection accuracy
  • Hard limits: Some models cap the number of tools they can handle

The Solution​

mcpzip sits between your AI assistant and your MCP servers. It replaces all those tool definitions with exactly 3:

Meta-ToolWhat It Does
search_toolsFind tools by keyword or natural language
describe_toolGet the full schema for a specific tool
execute_toolRun a tool on its upstream server

The model searches for what it needs, inspects the schema, and executes — loading only the tools it actually uses.

How It Works​

Claude Code                mcpzip                    MCP Servers
| | |
|-- search_tools ------->| |
| "send a message" |-- (keyword + LLM search) |
|<-- results ------------| |
| slack__send_message | |
| telegram__send_msg | |
| | |
|-- describe_tool ------>| |
| slack__send_message | |
|<-- full schema --------| |
| | |
|-- execute_tool ------->| |
| slack__send_message |-- tools/call ----------->|
| {channel, text} |<-- result ----------------|
|<-- result -------------| |

Key Features​

  • Context compression — 3 tools instead of hundreds
  • Smart search — keyword matching + optional Gemini semantic search
  • Instant startup — serves from disk cache, refreshes in background
  • All transports — stdio, HTTP (Streamable HTTP), SSE
  • OAuth 2.1 — browser-based PKCE flow, reuses mcp-remote tokens
  • Connection pooling — idle timeout, automatic reconnection
  • Per-call timeout — override defaults for slow tools
  • Auto-migration — import Claude Code MCP config in one command
  • ~5MB binary — single static binary, no runtime dependencies

Built by​

Hypercall