Cool project—Envoy dataplane + tiny routing LLMs is a solid combo for agent handoffs without the usual orchestration bloat.

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.

> We previously shared an early version of this project as ArchGW

Looks like these are the previous threads, if anyone's curious:

Show HN: ArchGW – An intelligent edge and service proxy for agents - https://news.ycombinator.com/item?id=44546265 - July 2025 (15 comments)

Show HN: ArchGW – An open-source intelligent proxy server for prompts - https://news.ycombinator.com/item?id=43259862 - March 2025 (7 comments)

Show HN: archgw: open-source, intelligent proxy for AI agents, built on Envoy - https://news.ycombinator.com/item?id=42187132 - Nov 2024 (14 comments)