Agentic observability with Amazon OpenSearch Service MCP Apps
Amazon Web Services Hang Zuo
AWS put observability charts inside the AI chat window. It cuts out the annoying tab-hopping between the agent and the dashboard.
Based on reporting by Amazon Web Services, Hang Zuo — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Amazon OpenSearch Service now has MCP Apps, a way to let observability agents answer with more than plain text. Instead of getting a root-cause guess and then wandering off to another browser tab to check it, the engineer can see the relevant chart, trace waterfall, or service map right there in the same conversation thread.
That matters because the slow part of agentic observability was never the first query. It was the verification loop. The agent could search alerts, compare logs with traces, and sketch a hypothesis in minutes. Then a human had to leave the IDE, open an observability UI, rerun the query by hand, and make sure the story on screen matched the story in the summary. MCP Apps are meant to remove that second trip.
The setup is built around a local MCP server that sits between the IDE and OpenSearch UI. The server runs on the user’s machine, forwards the request to an OpenSearch UI application, and returns two outputs at once: text for the agent and a visual widget for the person. AWS says the visualization is generated server-side from the same data sources behind the dashboards, so what shows up in the chat window is meant to match the actual query result, not the model’s best guess about it.
OpenSearch UI can connect to OpenSearch domains, serverless collections, CloudWatch, and Amazon Managed Service for Prometheus. The toolset spans triage, logs, traces, metrics, service maps, charts, datasets, correlations, LLM tracing, stack health, and instrumentation scoring. In practice, AWS is trying to collapse an investigation that used to bounce between tools into one thread where the agent investigates, the human verifies, and nobody loses their place.
Getting it running means downloading the MCP server, pointing an IDE such as Claude Desktop, VS Code GitHub Copilot, Goose, ChatGPT, or Cursor at the local server, and setting the OpenSearch UI endpoint plus AWS credentials. AWS also says the connection can be tested with a prompt asking for available observability data sources. If the answer comes back, the setup is working. If not, the IAM permissions need another look.
My take — AI-written commentary, not fact-checked reporting
This is the sort of feature that makes agentic tooling feel less like a demo and more like a real workflow. The industry keeps pretending that a good summary is the same as proof; it isn’t. Putting the evidence in the chat is the sensible move, which is probably why it took so long to show up.
Read more about this at: Amazon Web Services