Slim Mcp

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

Benchmarks

Tested on 57 tools across 4 real MCP servers. Accuracy validated with 120 API calls against Claude Sonnet 4.

Level	Tokens	Reduction	Accuracy
none	7,528	baseline	100%
standard	6,100	19%	100%
aggressive	4,930	35%	100%
extreme	2,133	72%	100%
maximum	1,750	77%	100%

With lazy loading (57 tools): 7,702 -> 2,722 tokens (65% reduction). Compression and lazy loading stack.

How extreme/maximum work: Instead of full JSON Schema, slim-mcp embeds TypeScript-style parameter signatures in the tool description and strips the inputSchema. The LLM reads the description to understand parameters -- which is what it does anyway.

Install

npm install -g slim-mcp

Quick Start

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}" }
    }
  },
  "compression": "extreme"
}

slim-mcp

Remote server:

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

Agent Integration

Claude Code

In .mcp.json:

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

Or wrap a single server:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "slim-mcp", "--", "npx", "-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
    }
  }
}

Any MCP Client

slim-mcp speaks standard MCP over stdio. Works with Cursor, Windsurf, or any client -- replace the server command with slim-mcp wrapping it.

Features

Feature	How it works	When it activates
Compression	5 levels from structural cleanup to TS-signature embedding	Always (default: standard)
Lazy loading	Slim one-liner indexes for most tools, full schema on first call	Auto when >15 tools
Caching	TTL + LRU for read-only calls, invalidates on writes	Always (disable with --no-cache)
Multi-server	Aggregates servers behind one proxy, namespaces as server__tool	Config file with 2+ servers
Remote transport	HTTP/SSE alongside local stdio servers	url in config or --url flag
Live dashboard	Real-time stats: compression, cache hits, tool calls, server status	Multi-server mode (port 7333)

Compression Levels

Level	What it does
none	Passthrough
standard	Structural cleanup, description trimming
aggressive	+ strips obvious descriptions, deduplicates params
extreme	Embeds TS-style signatures in descriptions, strips inputSchema (72% reduction)
maximum	Ultra-short types (s/n/b), ! for required, shared param extraction (77% reduction)

Dashboard

When running in multi-server mode, slim-mcp serves a live web dashboard:

http://localhost:7333

Shows: token savings, cache hit rate, server status, recent tool calls with HIT/MISS/PROMOTED status and response times. Updates in real-time via SSE. Dark theme, zero dependencies.

Enable via config:

{ "dashboard": { "enabled": true, "port": 7333 } }

Or CLI: --dashboard-port 7333. Disable: --no-dashboard.

Configuration

slim-mcp looks for .slim-mcp.json in the working directory, then home directory. Use --config <path> to specify explicitly.

{
  "servers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"],
      "always_load": ["read_file", "list_directory"],
      "cache_ttl": 30
    },
    "remote-api": {
      "url": "https://mcp.example.com/mcp",
      "type": "http",
      "headers": { "Authorization": "Bearer ${API_TOKEN}" }
    }
  },
  "compression": "extreme",
  "max_tools_loaded": 10,
  "cache": {
    "default_ttl": 60,
    "max_entries": 1000,
    "never_cache": ["my_custom_write_tool"]
  },
  "dashboard": {
    "enabled": true,
    "port": 7333
  }
}

Each server needs command (stdio) or url (HTTP/SSE), not both. Environment variables expand with ${VAR} syntax.

Full reference in docs/configuration.md.

CLI

slim-mcp [options] -- <command> [args...]     Single server (stdio)
slim-mcp [options] --url <url>                Single server (HTTP/SSE)
slim-mcp [options] --config <path>            Multi-server (config file)
slim-mcp [options]                            Multi-server (auto-discover)

Flag	Default
-v, --verbose	off	Show JSON-RPC messages on stderr
--compression <level>	standard	none, standard, aggressive, extreme, maximum
--no-cache		Disable response caching
--no-lazy		Disable lazy loading
--max-tools <N>	8	Max tools with full schemas
--url <url>		Remote MCP server URL
--header <key:value>		HTTP header (repeatable)
--transport <type>	auto	http or sse
--dashboard-port <N>	7333	Enable dashboard on port
--no-dashboard		Disable dashboard
--config <path>		Config file path
--version		Show version

How It Works

                                ┌── Server A (stdio)
Agent <--stdio--> slim-mcp ----+-- Server B (stdio)
                                └── Server C (http)

On tools/list: collect from all servers -> lazy load (slim/full split) -> compress -> return.

On tools/call: check if slim (promote + retry) -> check cache (return if hit) -> route to server -> cache result.

All logging to stderr. Protocol channel stays clean.

Architecture deep dive in docs/how-it-works.md.

Testing

220 tests: 190 unit + 30 e2e. Plus accuracy tests and smoke tests against real servers.

npm test              # Unit tests
npm run test:e2e      # E2E tests
npm run smoke-test    # Real servers from ~/.claude.json

Accuracy Testing

Validates that compressed schemas produce correct tool calls via the Anthropic API.

ANTHROPIC_API_KEY=sk-... npx tsx scripts/accuracy-test.ts

• 8 test scenarios x 5 compression levels x 3 runs = 120 API calls
• Validates: tool selection, argument names, argument types
• Cost: ~$0.20 per run
• Result: 100% accuracy across all levels

Details in docs/testing.md.

Requirements

• Node.js >= 18

License

MIT