slim-mcp

MCP proxy that compresses tool schemas, lazy-loads definitions, caches responses, and aggregates multiple servers -- giving AI agents their context window back.

Tested against 3 real MCP servers (42 tools total) on a production Claude Code machine:

| Feature | What happened | Tokens | |---------|--------------|--------| | Without slim-mcp | 42 raw tool schemas | 5,812 | | Lazy loading | 42 slim indexes | 2,191 (62% saved) | | Compression | Trimmed descriptions, deduped params | Stacks on full schemas | | Caching | 2nd identical call → instant | ~1,644 saved per repeated call |

Single server (8 tools): 715 → 611 tokens (15% compression). Multi-server (42 tools): lazy loading dominates at 62%. Both features stack -- a 50-tool setup with full schemas sees 80-95% combined reduction.

Every tool call routed correctly. Cache hit confirmed on repeated calls. All 203 tests pass.

npm install -g slim-mcp

Single server -- wrap any MCP server:

slim-mcp -- npx -y @modelcontextprotocol/server-filesystem /tmp

Multiple servers -- create .slim-mcp.json:

{
  "servers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}" }
    }
  }
}

slim-mcp

Remote server:

slim-mcp --url https://mcp.example.com/mcp --header "Authorization:Bearer $TOKEN"

In .mcp.json:

{
  "mcpServers": {
    "tools": {
      "command": "npx",
      "args": ["slim-mcp", "--config", "/path/to/.slim-mcp.json"]
    }
  }
}

Or wrap a single server:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["slim-mcp", "--", "npx", "-y", "@mode