# MCPs for Observability Stacks

Diana Todea, VictoriaMetrics | MLOps Community | 24:28

Source: https://www.youtube.com/watch?v=aQHg6db9wRs
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/mcps-for-observability-stacks
Published: 2026-08-19
Tags: human-in-the-loop, mcp, monitoring, observability

## TL;DR
- MCP servers can connect an AI assistant to observability data, APIs, documentation, and operational tools.
- Anomaly detection can replace static thresholds with model-based detection of unusual behavior, while business boundaries provide context for interpreting results.
- Two MCP servers can work together, with one querying VictoriaMetrics data and another selecting, validating, and configuring anomaly detection models and alerts.

## Summary
Diana Todea shows how Model Context Protocol can connect an AI assistant to an observability stack. Her examples combine a VictoriaMetrics cloud deployment with a separate anomaly detection MCP server. The assistant retrieves metrics and logs, reviews metric usage, searches documentation, selects anomaly detection models, and drafts alert configurations. Todea first explains the value of bringing metrics, logs, traces, and events into one system context. She then demonstrates anomaly detection on log-derived error-span data, including prediction bands, anomaly boundaries, and model configuration. In the main workflow, the assistant provides a deployment overview, identifies unused metrics, creates dashboard information, recommends a model, and produces a draft YAML alert. A second example profiles a latency query and recommends changing the model from Prophet to an online model. Todea keeps a human approval step before configuration changes are applied. She also says that standards for MCP and semantic conventions are still developing, and that real-world validation remains necessary.

## Key ideas
### Observability work needs system context before automation
[01:30](https://www.youtube.com/watch?v=aQHg6db9wRs&t=90s)
Diana says developers want metrics, logs, traces, and events brought together so they can understand the full system. They also want early anomaly detection and automation for repetitive operational work. During an incident, the aim is to reduce mean time to resolution and speed up root cause analysis. Correlated telemetry should make troubleshooting faster, while reliable and available data gives engineers a basis for decisions.

### MCP connects an AI agent to live observability tools
[02:40](https://www.youtube.com/watch?v=aQHg6db9wRs&t=160s)
MCP can query and correlate telemetry across an observability stack. It can also let an agent investigate anomalies, retrieve metrics, and guide root cause analysis. Diana contrasts this with manually moving through dashboards and writing complex queries. Natural language gives the assistant a way to access tools, APIs, documentation, and data, while still allowing it to carry out actions rather than only describe them.

### Anomaly detection identifies unusual patterns against expected behavior
[03:43](https://www.youtube.com/watch?v=aQHg6db9wRs&t=223s)
Diana defines an anomaly as a data point or pattern that differs significantly from expected behavior. Spikes, drops, data-collection errors, equipment problems, sudden behavior changes, and data drift can all produce anomalies. In her example, the model predicts a normal range for error-span data and marks unusual points with red dots. She also uses direction and minimum deviation settings to focus on the anomalies that matter.

### Business boundaries give anomaly results operational context
[07:23](https://www.youtube.com/watch?v=aQHg6db9wRs&t=443s)
The anomaly view becomes more useful when business boundaries are configured. Without those boundaries, spikes and drops can be technically unusual without being meaningful for the service. Diana shows that the boundaries help specify what engineers are looking for and make the anomaly output easier to interpret. The model configuration can be exported, and an example alert can be created for review and sharing with a team.

### Skills package repeatable troubleshooting steps for an agent
[09:29](https://www.youtube.com/watch?v=aQHg6db9wRs&t=569s)
Diana describes MCP skills as reusable playbooks, recipes, or user manuals. They contain recommended steps, guardrails, and examples for a particular task. Her setup includes skills for querying, reviewing anomaly detection, configuring models, and working with alerting. The skills tell the agent how to perform a task while the MCP servers provide access to observability data and documentation.

### Two MCP servers divide observability access from anomaly analysis
[12:24](https://www.youtube.com/watch?v=aQHg6db9wRs&t=744s)
The demonstration connects two MCP servers to Claude Desktop. One provides access to a VictoriaMetrics cloud observability deployment. The other provides the anomaly detection backend. Diana recommends read-only tools for this kind of setup. The first server returns a deployment overview with scrape information, storage, ingestion rates, resource usage, health, and a metric catalogue. The second server handles model selection and configuration.

### The assistant can find unused metrics and draft a dashboard
[14:43](https://www.youtube.com/watch?v=aQHg6db9wRs&t=883s)
Diana asks the observability MCP server to identify frequently used and unused metrics. In her example, 98% of the metrics had not been used before, which indicated that cleanup was needed. The assistant can also provide information for a more focused dashboard, including the proportions of used and unused metrics. This lets an engineer inspect a deployment before an incident rather than waiting for a failure.

### Model changes remain subject to human approval
[20:12](https://www.youtube.com/watch?v=aQHg6db9wRs&t=1212s)
In the anomaly workflow, the assistant profiles a query's time-series characteristics, checks available models and schemas, and validates a proposed configuration. The profile has no trend or daily or weekly seasonality and is a poor match for the current model. The assistant recommends switching to an online model. Diana checks the recommendation before applying it, and the configuration changes only after she confirms that it is reasonable.

## Notable quotes
- "We want to bring the metrics, logs, traces, events together and be able to get out a complete system context out of it." (01:31)
- "MCP is very good at connecting an AI agent to live tools with APIs, with documentation and with data." (09:08)
- "They are reusable playbooks that will teach an AI agent how to accomplish a particular object." (09:30)
- "I always recommend to use the readonly tools." (12:48)
- "We are the human in the loop here. It doesn't go ahead and apply the configuration change without us verifying this." (21:02)

## Tools & references mentioned
- VictoriaMetrics
- Model Context Protocol
- MCP
- Claude Desktop
- Cursor
- Anthropic
- OpenTelemetry
- semantic conventions
- YAML
- alerting

## Who should watch
- You operate production systems and want an assistant to query metrics, logs, and documentation without making engineers move through several observability interfaces.
- Your team is testing anomaly detection and needs a workflow for model selection, configuration, and alert review.
- You are considering MCP for operational tooling and want an example that keeps read-only access and human approval in the workflow.

## Related talks

- [Enterprise-ready MCP](https://mlopstalks.com/talks/enterprise-ready-mcp) (Jiquan Ngiam, MintMCP, 28:36)
- [MCP Security: What Happens When Your Agents Talk to Everything?](https://mlopstalks.com/talks/mcp-security-what-happens-when-your-agents-talk-to-everything) (, 24:26)
- [Yes, we do need MCP](https://mlopstalks.com/talks/yes-we-do-need-mcp) (Ben Hindman, reboot.dev, 26:32)
- [MCP is not going to change everything (yet)](https://mlopstalks.com/talks/mcp-is-not-going-to-change-everything-yet) (Sam Partee, Arcade AI & Rahul Parundekar, AI Hero, 1:04:43)
- [Prepare Your Data for AI Now, or You're Screwed Later](https://mlopstalks.com/talks/prepare-your-data-for-ai-now-or-youre-screwed-later) (Simba Khadder, Featureform, 22:04)