An MCP proxy server that reduces Claude Code's context window usage by keeping tool definitions server-side and using AWS Bedrock's tool search to find relevant tools on-demand.

With 50+ tools from multiple MCP servers, Claude Code's context becomes bloated (~10,000 tokens). This proxy keeps tools server-side and only exposes 2 tools to Claude Code.

Claude Code → Tool Proxy (2 tools in context)
                   |
                   ├── search_tools(query) → Bedrock tool search → returns tool names
                   ├── call_tool(name, args) → routes to backend
                   |
                   └── Backend MCP servers (tools cached server-side)

• Node.js 20+
• AWS account with Bedrock access
• Claude Opus 4.5 model enabled in Bedrock
• AWS credentials configured

npm install
npm run build

Configure backend MCP servers to proxy:

{
  "backends": {
    "aws-knowledge": {
      "url": "https://knowledge-mcp.global.api.aws",
      "type": "http"
    },
    "github": {
      "url": "https://api.githubcopilot.com/mcp/",
      "type": "http",
      "headers": {
        "Authorization": "Bearer ${GITHUB_PAT}"
      }
    }
  }
}

Configure Claude Code to connect only to the proxy:

{
  "mcpServers": {
    "tool-proxy": {
      "command": "npx",
      "args": ["-y", "tool-search-bedrock-mcp"]
    }
  }
}

The proxy exposes 2 tools:

Search for relevant tools by query:

{
  "query": "I need to search AWS documentation",
  "max_results": 5
}

Returns tool names that can be used with call_tool.

Execute a discovered tool:

{
  "tool_name": "aws-knowledge__aws___search_documentation",
  "arguments": {
    "search_phrase": "Lambda best practices"
  }
}