paidev
Joined 10 months ago
Joined 10 months ago
We've wrestled with similar delivery pains building MCP tools for agents (Claude/ChatGPT/Cursor all love 'em now). Proxies like yours shine for LLM routing, but tool backends often drag with auth/setup. Your MCP integration in filters caught my eye—@leanmcp/auth decorator drops proper JWT validation (Cognito/Auth0/etc.) to 20 lines vs 600+ raw, auto-injects user context everywhere. (Disclosure: co-founder on LeanMCP.)
How's the Brightstaff fallback to static policies holding up in prod? Happy to chat agent war stories.
On the auth side, GitHub/Google OAuth is smart for bumping limits, but raw MCP makes JWT validation a slog (we burned days on it early). If you're open-sourcing the impl, `@leanmcp/auth` decorator drops it to ~20 lines with auto-injected user context—handles multiple providers out of the box.
How'd you wire up the rate limiting backend to tie logins to IPs without killing perf? (Disclosure: co-founder at LeanMCP)