MCP Slim

Your MCP servers eat 72% of your context window. Fix it in 60 seconds.

Connect 3 MCP servers and 55,000 tokens are gone before your first message. That's 27% of Claude's context window. On tool definitions you'll never use.

MCP Slim sits between your client and your servers. Your AI sees 3 tools instead of 100+.

How it works

search_tools("create github issue")     →  5 matches, ~200 tokens
get_tool_schema("github_create_issue")  →  just that one schema
call_tool("github_create_issue", {...}) →  routed to the right backend

That's it. Instead of loading every tool upfront, the LLM searches for what it needs, looks up the schema, and calls it. MCP Slim handles the routing.

┌─────────────┐         ┌──────────────────────┐
│  Claude /    │  stdio  │      MCP Slim        │
│  Cursor /    │◀──────▶│                      │
│  VS Code     │         │  search_tools        │
└─────────────┘         │  get_tool_schema     │
                         │  call_tool           │
                         └───┬──────┬──────┬───┘
                             │      │      │
                          GitHub  Slack  Sentry  ...

Zero changes to your servers. Zero changes to your client.

Savings

| Layer | Before | After | Reduction | |-------|--------|-------|-----------| | Tool catalog | 20,000 tokens | 270 tokens | 98.7% | | Schema per lookup | 300 tokens | 180 tokens | 39% | | API responses | 40,000 tokens | 5,000 tokens | 87% |

On a typical 3-tool task: ~20,000 tokens down to ~700. That's 96%+.

Install

npx mcp-slim init

Detects your existing Claude Desktop or Cursor config, backs it up, and rewrites it. Done.

Note: First run downloads a ~80MB embedding model. Search works immediately via keyword matching; semantic search activates once the model loads (~2 seconds).

Or configure manually

// Your MCP client config
{
  "mcpServers": {
    "mcp-slim": {
      "command": "npx",
      "args": ["mcp-slim", "proxy"]
    }
  }
}

// ~/.mcp-slim/config.json
{
  "backends": {
    "github": {
      "command": "github-mcp-server",
      "env": { "GITHUB_TOKEN": "ghp_..." }
    },
    "slack": {
      "command": "slack-mcp-server",
      "args": ["--workspace", "myteam"]
    }
  }
}

What it does under the hood

• Hybrid search — keyword matching + local semantic embeddings (@huggingface/transformers). "Save a note" finds create_memory_entity even with zero keyword overlap.
• Schema compression — strips redundant descriptions, type-default values, JSON Schema meta-fields.
• Response optimization — truncates arrays, strips metadata keys by pattern, removes nulls.
• Usage tracking — per-request token savings logged to stderr, session totals persisted to SQLite. Shows estimated cost savings on shutdown.
• Async model loading — fuzzy search works instantly, embedding search kicks in once the model loads (~2 seconds).
• Graceful degradation — if embeddings fail (offline, low memory), falls back to fuzzy. Never crashes the proxy.

CLI

mcp-slim proxy              # Start the proxy (run by MCP clients)
mcp-slim proxy --verbose    # Debug logging to stderr
mcp-slim init               # Auto-configure from existing client config
mcp-slim status             # Show backends, settings, all-time token savings

Config

All optional. Defaults work out of the box.

{
  "backends": {},
  "maxToolsPerSearch": 5,
  "searchMode": "hybrid",
  "embeddingModel": "Xenova/all-MiniLM-L6-v2",
  "enableSchemaCompression": true,
  "enableResponseCompression": true,
  "maxArrayItems": 10,
  "maxStringLength": 2000,
  "stripKeyPatterns": ["_links", "_meta", "node_id", "*_url"],
  "removeNulls": true,
  "trackUsage": true
}

Works with

Any MCP client (Claude Desktop, Cursor, VS Code, custom agents) and any MCP server. stdio transport.

License

MIT