MCP Shrink: your MCP servers eat a third of your context window before you even type
We measured 7 MCP servers consuming 67,300 tokens — 33.7% of a 200k window — before the first message. The open-source compressors all run locally as stdio proxies: clone, configure, run Node, re-point every client, on every machine. Nobody offers a hosted one.
What it would do
- Point your MCP client at one URL:
https://<namespace>.updatesbyai.com/mcp/<upstream-url> - Tool definitions compressed to ~54-token stubs, lazy-loaded only when called
- Tool responses distilled before they hit your context window
- Live dashboard: tokens saved per session, per server
- Works with Claude Desktop, Cursor, Cline, Claude Code — any MCP client
Planned pricing: free for 3 servers, $9/mo unlimited. We're validating before building — waitlist size decides whether this gets built at all.
Join the waitlist
No spam, no launch emails — one message when it's live (and maybe one question about your setup).