Anyone else running into context limits with MCP tool catalogs?
Tried the Apideck MCP Server this week, which solves a real problem for agent builders
Been experimenting with MCP servers for connecting agents t business data and ran into the usual issues: bloated tool lists, auth complexity, and context windows getting hammered.
Apideck just launched one that handles this differently. Single MCP server, 200+ connectors (QuickBooks, Xero, NetSuite, and others), normalized data across all of them.
Two things stood out:
- Dynamic context mode loads 4 meta-tools instead of the full catalog. You discover and call tools on demand. They claim significant context reduction, around 63x by their numbers, which roughly matched what I saw.
- Scoped permissions let you restrict what the agent can even see. Useful if you want read-only access to start before it writes anything.
Works with Claude, ChatGPT, Cursor, LangChain, and any other tool that supports MCP.
If you're building agents that need to pull from actual business systems rather than toy data, worth a look. They launched on Product Hunt today: https://www.producthunt.com/products/apideck/launches/apideck-mcp-server
Anyone else been wrestling with the context bloat problem on MCP? Curious what approaches you've tried.