Post

Slowly migrating to this at work I.e people oncall rarely open the ui and instead go llm -> provider mcp…. So if the data is in the write shape to query then why pay $$ to the o11y provider 🐶

Want to debug something? Ask an agent Need a graph of something? Ask an agent Need a dashboard? Ask an agent

You may still need the data to be ingested into somewhere. Ideally in the same format. Then the ask the agent can be just as easy. Both for you and the agent. 😎 But if it's a bunch of disjointed data points. Scattered like dragon balls... No amount of context window will make that better. 🫠

Have you tried pointing an agent at ClickHouse? It’s an… informative experience

Think that approach works for some people but can still see cases where people want pre-bake visualisations as they don’t know what’s possible or don’t want to wait to an agent to figure out what it needs to do But yeh agents and telemetry are a good fit

The above is a slight over-exaggeration But I’m less and less sure by how much

Kinda? I agree the UI part isn't the "most" useful part. Where the UI becomes useful is helping a human understand the data, and claude artifacts are "fine", but still not great. From a backend perspective, its more than just querying data with SQL/PromQL. The backend can do more.

We've been seeing this with people resorting to "codemode" type mechanisms. We've taken a different stance and our MCP does more than just an API reference, it does so much more than that, and does it server-side. Providing additional context, and rendering is something the backend can do.

I don't think MCP should be just an API reference so I agree there, but also I wouldn't call codemode something to "resort" to. It's a really important primitive. I'm also trying to think a year+ out. Claude artifacts are not great... yet.