Refract
MCP proxy that compresses tool schemas on the fly. Up to 98% token reduction, 100% signal preserved verified after every compression. Zero LLM calls, fully deterministic.
README
<p align="center"> <img src="assets/logo.png" alt="Refract" width="360"> </p>
Refract
Cuts up to 97% of the tokens your AI agents spend using MCP tools — without losing anything.
The problem in one sentence
When your AI agent (Claude, Cursor...) connects to an external tool — your calendar, your emails, GitHub — it downloads the full description of every available tool, every single time, even if it only ends up using one.
It's like asking someone to read the entire store catalogue just to buy bread.
Refract fixes that. It sits between your agent and the tool server, and only lets through what's actually needed.
What it actually changes
| Without Refract | With Refract | |
|---|---|---|
| Filesystem tools (14 tools) | 1,892 tokens | 236 tokens (−88%) |
| Google Calendar tools (5 tools) | 5,010 tokens | 660 tokens (−87%) |
| Enterprise pack — Calendar + Gmail + Drive (12 tools) | 8,649 tokens | 882 tokens (−90%) |
Fewer tokens sent = lower API bills, faster responses.
And nothing is lost. Every check confirmed tools stay 100% usable after compression and no required information is ever stripped.
Install
pip install refract-mcp
That's it. No API key required, no account needed.
How to use it
With Claude Desktop
Open your Claude Desktop config file and add:
{
"mcpServers": {
"my-tool-via-refract": {
"command": "refract-proxy",
"args": [
"--target",
"npx @modelcontextprotocol/server-filesystem /path/to/folder",
"--verbose"
]
}
}
}
Replace the --target line with any MCP server you already use. Restart Claude Desktop and that's it, Refract runs in the background.
You may have this message when launching Claude Desktop :"Failed to spawn process: No such file or directory" in Claude Desktop
This usually means Claude Desktop can't find refract-proxy in its PATH. Find the absolute path and use it directly in your config:
which refract-proxy
Then use the full path in claude_desktop_config.json:
{
"mcpServers": {
"my-tool-via-refract": {
"command": "/full/path/to/refract-proxy",
"args": [
"--target",
"npx @modelcontextprotocol/server-filesystem /path/to/folder",
"--verbose"
]
}
}
}
From the command line
refract-proxy --target "npx @modelcontextprotocol/server-filesystem /tmp" --verbose
The --verbose flag shows live savings:
[Refract] Connected to npx @modelcontextprotocol/server-filesystem /tmp
14 tools | 1892 → 236 tokens (88% reduction)
How it works, no jargon
Refract optimizes information retrieval by sending an agent a simple index of tool names instead of a massive summary of all available data. Once the required tool is identified, the system delivers only the necessary full details and automatically verifies that no important content was lost during the compression. Operating completely without artificial intelligence, this streamlined process is entirely automated, fast, and predictable.
Works with
- Claude Desktop
- Cursor
- Any client that follows the MCP (Model Context Protocol) standard
- Any existing MCP server — your internal tools, GitHub, Google Workspace, Slack, etc.
For developers
Python usage
from refract_proxy import RefractProxy
proxy = RefractProxy(
target_url="npx @modelcontextprotocol/server-filesystem /tmp",
verbose=True,
)
await proxy.connect()
# Use compressed tools directly with the Anthropic API
tools = proxy.as_anthropic_tools(use_cache=True)
# Or serve as a local MCP server (stdio)
await proxy.serve()
# Or expose it via HTTP/SSE
await proxy.serve_http() # → http://localhost:8080/sse
HTTP/SSE mode
refract-proxy --target "https://my-mcp-server.com" --mode http --port 8080
With a local schema file (for testing)
refract-proxy --target schemas/mcp_calendar_schemas.json --verbose
Built-in Anthropic caching
Refract integrates with Anthropic prompt caching: as_anthropic_tools() automatically marks the compressed catalogue as cacheable, cutting costs even further on repeated requests.
Example over 30 days, 100 requests/day, 5,000 tokens of schemas:
| Cost | |
|---|---|
| Without Refract, without cache | $45.00 |
| With Refract + cache | $1.49 |
License
MIT — free to use, including commercially. ENDOFREADME
cd ~/MCP-repo git add README.md git commit -m "Force English README" git push origin main
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。