It always starts with “just one integration”.
You want your AI agent to send a message to Slack. So you wire it up. A bit of custom code, some API...
For further actions, you may consider blocking this person and/or reporting abuse
Approval workflows catch the operations you anticipated. They don't catch an agent looping on the same tool because its reasoning got stuck. The gateway is also the right place for cost caps per run and circuit breakers per tool, otherwise the first surprise bill teaches you the lesson.
Exactly!
Approval flows only cover what you expect; the real pain is when agents go off-script. Cost caps and circuit breakers at the gateway level honestly feel non-negotiable once you’ve seen it happen.
Great breakdown of MCP beyond the hype. I especially liked how you highlighted the gap between protocol standardization and production readiness.
Many teams assume MCP is “the solution”, while in reality it only solves the communication layer. The need for a gateway for access control, observability, and security becomes obvious as soon as systems scale.
Really glad that landed; that gap is exactly what I wanted to highlight.
MCP gets you a clean way to connect things, but scaling is where the need for control and visibility hits hard.
MCP is great until you actually try to run it in production.
The “it solves communication, not control” part hit hard. That’s exactly where things start breaking, and nobody talks about it.
Really solid 👏🏻
Glad that part resonated; that’s exactly the gap I kept running into too.
MCP makes things work, but production is where you realize how much is still missing around control and visibility.
Appreciate you reading it 😍
Great breakdown of why MCP needs a gateway layer. The raw protocol is powerful but in production you definitely need auth, rate limiting, and observability between the client and the tool servers. We have been running MCP servers internally and the biggest gap we hit was exactly this - no standardized way to handle multi-tenant access control. Your nginx/Envoy approach makes a lot of sense for teams already using service mesh. For smaller setups, a simple Node.js proxy with JWT validation in front of the MCP server works well - lighter weight but still gives you the auth boundary.
Super practical take.
Multi-tenant access control is exactly where things get messy fast. And yeah, totally agree: you don’t always need a heavy setup. A simple Node proxy with JWT can go a long way early on.
Love the service mesh angle too; that’s a natural fit for bigger systems.
Great breakdown.
Thank you! Glad you found it helpful.
Solving integrations is one step, running them reliably in production is the real challenge.
Exactly! That’s the part that usually gets underestimated.
Integrations are easy to demo, but production is where all the edge cases, failures, and scaling issues show up.