DEV Community

AyhanYetgin
AyhanYetgin

Posted on • Originally published at contextopt.vercel.app

I Built an Open Source Tool to Cut Claude Code Token Waste by 70%"

#claude #mcp #opensource #typescript

What is ContextOpt?

ContextOpt is a smart proxy that sits between your AI agent and its MCP servers. It filters tools by profile so only relevant ones consume context.

AI Agent (Claude Code)
       │
       ▼
┌─────────────────┐
│  ContextOpt      │  ← Filters tools by profile
│  Proxy Engine    │  ← Lazily starts servers
└────────┬────────┘
         │
    ┌────┴────┐
    ▼         ▼
filesystem  github  ... (only active profile servers)
Enter fullscreen mode Exit fullscreen mode

The Problem

Here's a typical MCP config for Claude Code:

{
  "mcpServers": {
    "filesystem": { "command": "npx", "args": ["-y", "@anthropic/mcp-filesystem"] },
    "github": { "command": "npx", "args": ["-y", "@modelcontextprotocol/github"] },
    "playwright": { "command": "npx", "args": ["-y", "@anthropic/mcp-playwright"] },
    "sequential-thinking": { "command": "npx", "args": ["-y", "@anthropic/mcp-sequential-thinking"] },
    "memory": { "command": "npx", "args": ["-y", "@anthropic/mcp-memory"] }
  }
}
Enter fullscreen mode Exit fullscreen mode

That's 5 servers → ~78 tool definitions → ~19,500 tokens per prompt. Every. Single. Message.

If you send 50 messages a day, that's nearly 1M tokens/month just on tool definitions — about $90/month on Claude Sonnet.

The Solution: Profile-Based Filtering

Instead of loading every tool all the time, ContextOpt groups servers into profiles:

Profile Servers Active Tokens Savings
default (all) 5 servers ~19.5K baseline
coding filesystem, github ~10K 49%
debugging playwright, memory ~8.8K 55%
research sequential-thinking, memory ~2K 90%

When you're coding, you don't need Playwright. When you're debugging, you don't need GitHub. Simple.

How It Works

1. Analyze your current setup 

npx contextopt analyze
Enter fullscreen mode Exit fullscreen mode

This reads your MCP config (auto-detects Claude Code, Cursor, or Windsurf) and shows a breakdown of token usage per server, estimated cost, and potential savings.

2. Pick a profile 

# List available profiles
npx contextopt profile -l

# Switch to coding profile
npx contextopt profile -s coding

# Create a custom profile
npx contextopt profile -c myprofile --servers "github,filesystem"
Enter fullscreen mode Exit fullscreen mode

3. Start the proxy 

npx contextopt start --profile coding
Enter fullscreen mode Exit fullscreen mode

The proxy connects to only the servers in your active profile, discovers their tools, and exposes the filtered list to your AI agent. Servers you don't need are never started — saving both tokens and system resources.

HTTP Mode (Pro)

For remote connections, you can run the proxy in HTTP mode:

npx contextopt start --http --port 3456
Enter fullscreen mode Exit fullscreen mode

This uses Streamable HTTP transport, so any MCP-compatible client can connect.

The Web Dashboard

You don't have to use the CLI if you prefer a visual interface. The live dashboard lets you:

  • Paste your MCP config and see instant token analysis
  • Compare savings across profiles
  • See per-server breakdown with charts
  • Calculate your monthly savings

Dashboard

What I Learned Building This

This was my first time working with the MCP SDK. A few things I learned:

Server lifecycle management is tricky. MCP servers are spawned as child processes, and you need to handle startup, keepalive, and graceful shutdown. The SDK handles most of this, but I had to build a wrapper around it for lazy loading.

Token estimation is never exact. I use a heuristic based on typical tool counts per server (filesystem ~15 tools, github ~25, etc.) and average tool schema size (~250 tokens/tool). It's close enough for planning, but actual usage varies.

The proxy pattern is powerful. By acting as a transparent proxy, ContextOpt works with any MCP client without modification. The agent connects to ContextOpt, and ContextOpt connects to the real servers. No config changes needed.

Try It 

npx contextopt analyze
Enter fullscreen mode Exit fullscreen mode

No installation needed — npx handles everything.

Roadmap

  • [x] CLI with analyze, profile, start commands
  • [x] Profile system with persistent storage
  • [x] Proxy engine with lazy tool discovery
  • [x] Web dashboard with live token analytics
  • [x] HTTP transport for remote connections
  • [x] Pro subscription with Lemon Squeezy payments

Up next: usage analytics, team profiles, and a VS Code extension.

*ContextOpt is open source (MIT). If you find it useful, consider sponsoring the project or starring it on GitHub.

Top comments (0)