⚡ MCP Token Gateway

Cut 40-90% of MCP tool definition token overhead.

Dashboard · API Docs · Quick Start

---

The Problem

MCP tool definitions burn 75,000+ tokens before your agent writes a single line of code:

10 MCP servers × 15 tools × 500 tokens = 75,000 tokens per session
= $0.375 per conversation start (Claude Sonnet)
= $375/month for a 5-person team

The Solution

MCP Token Gateway is a lightweight proxy that compresses tool definitions on the fly:

• 40-50% savings with rules-based compression (free tier)
• Zero code changes — just change the MCP server URL
• <10ms latency — Cloudflare edge, 300+ cities worldwide
• Free tier — 1,000 calls/day, no credit card

Quick Start (30 Seconds)

1. Get your free API key

curl -X POST https://gateway.toolboxlite.com/api/register \
  -H "Content-Type: application/json" \
  -d '{"email": "you@example.com", "password": "yourpassword"}'

2. Update your MCP client config

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "github": {
      "url": "https://gateway.toolboxlite.com/sse/mtg_YOUR_KEY/https%3A%2F%2Fmcp.github.com%2Fsse"
    }
  }
}

Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "my-server": {
      "url": "https://gateway.toolboxlite.com/sse/mtg_YOUR_KEY/https%3A%2F%2Fyour-mcp-server.com%2Fsse"
    }
  }
}

3. Done! 🎉

Tool definitions are compressed automatically. Check your savings at the dashboard.

How It Works

Your AI Client  →  MCP Token Gateway  →  Downstream MCP Server
(Claude/Cursor)    (Cloudflare Edge)      (GitHub/Slack/etc)
                        ↓
              Compresses tools/list
              response on the fly
              (40-90% fewer tokens)

Compression Rules (V0.1)

1. Description trimming — Remove filler phrases ("This tool allows you to..."), keep first 2 sentences
2. Schema minification — Remove redundant property descriptions, examples, self-evident metadata
3. Whitespace collapse — Eliminate extra spaces, newlines, formatting
4. Enum description removal — If enum values are self-explanatory, drop the description

URL Format

https://gateway.toolboxlite.com/{transport}/{apiKey}/{encoded-target-url}

Parameter	Description
transport	proxy (HTTP POST) or sse (SSE)
apiKey	Your API key (starts with mtg_)
encoded-target-url	URL-encoded downstream MCP server URL

API Reference

Endpoint	Method	Description
/proxy/{key}/{url}	POST	Proxy MCP JSON-RPC request
/sse/{key}/{url}	GET/POST	Proxy MCP SSE connection
/api/register	POST	Create account ({email, password})
/api/login	POST	Login ({email, password})
/api/usage/{key}	GET	Get usage stats (?days=7)
/api/test-compress	POST	Test compression ({tools: [...]})
/health	GET	Health check

Dashboard

Visit gateway.toolboxlite.com/dashboard to:

• Register and manage API keys
• View token savings (before vs after)
• Monitor daily usage

Pricing

Plan	Price	Calls/day	Savings
Free	$0	1,000	40-50% (rules compression)
Developer	$19/mo	50,000	90%+ (full optimization)
Team	$49/mo	200,000	90%+ + team features

Known Limitations

Cloudflare-Proxied Downstream Servers (Error 1016)

If the downstream MCP server you're proxying is also behind Cloudflare's CDN proxy (orange-clouded), Cloudflare Workers will return a 1016 error due to CDN loop detection. This is a Cloudflare platform restriction.

Workarounds:

• Use the origin IP of the downstream server (bypass its Cloudflare proxy)
• Use a downstream URL that is not proxied through Cloudflare (DNS-only / gray-clouded)
• Self-host the gateway worker on a non-Cloudflare platform

This does not affect downstream servers hosted on other platforms (AWS, GCP, Vercel, Railway, etc.).

Architecture

• Proxy Layer: Cloudflare Workers (0ms cold start, global edge)
• Storage: Workers KV (API keys, usage stats)
• Dashboard: Next.js + Tailwind CSS on Vercel
• Cost: ~$0 at free tier volume

Self-Hosting

Worker

cd worker
npm install
cp wrangler.toml.example wrangler.toml  # Add your KV namespace IDs
npx wrangler deploy

Dashboard

cd dashboard
npm install
npm run build
# Deploy to Vercel or any Node.js host

Roadmap

• V0.1: Rules-based compression (40-50% savings)
• V0.2: Semantic tool discovery (90%+ savings)
• V0.3: Result caching
• V0.4: Team workspaces
• V0.5: Custom compression rules

License

MIT © Vincentwei1021