Best MCP Servers for Observability in 2026

The best MCP servers for observability in 2026 are Datadog, Grafana, Sentry, PagerDuty, New Relic, Honeycomb, Dynatrace, Elastic and Splunk. Each one lets an AI assistant read your logs, metrics, traces or incidents, so an engineer can ask "Why did checkout latency spike at 2 a.m.?" and get an answer from real telemetry.

MCP (Model Context Protocol) is the open standard that connects AI apps such as Claude, ChatGPT, Cursor and Codex to outside tools. Most observability vendors now host their own MCP server, which means no local install and sign-in through your existing account.

We checked every entry against the vendor's own site or docs on 28 October 2026. We did not test these products.

This list is about monitoring your own systems with AI. If you want to monitor the MCP servers and agents themselves, see MCP audit logging, learn MCP logging and monitoring and the best MCP observability tools.

How we chose

We used the MCP Guide's standard listicle criteria:

• Official. The vendor publishes and documents its own MCP server. Community servers were left out.
• Live today. We opened each vendor's docs and pricing page on 28 October 2026.
• Clear access model. The docs explain how the server signs in and what it can change.

Sentry also appears on our best MCP servers for developers list. Here it is covered for production error monitoring. The order below is not a ranking. More lists live on the MCP Guide's best-of hub.

Comparison at a glance

1. Datadog: logs, metrics, traces and monitors

What it is. The Datadog MCP Server lets an agent search logs, metrics, traces, hosts and incidents. Datadog says it "forwards the authenticated user's own credentials to Datadog APIs," and write tools such as creating monitors need the mcp_write permission. You can "limit available tools at connection time with toolsets and omit_tools."

Best for: Teams already on Datadog who want one AI entry point to all their telemetry.

Pricing: No separate MCP price. Datadog's pricing page lists Infrastructure Pro "Starting At $15 Per host, per month," billed annually. The docs list fair-use limits of "100,000 monthly tool calls."

Pros:

• Hosted, with OAuth recommended.
• Toolsets and omit_tools keep the tool list small.
• Works with many clients, including Claude Code, Cursor and Codex.

Cons:

• Datadog says the server "is under significant development."
• Not supported on GovCloud sites.

Checked against: Datadog docs on 28 October 2026. Full profile: Datadog MCP server.

2. Grafana: open-source access to your dashboards and data sources

What it is. The Grafana MCP server "provides access to your Grafana instance and the surrounding ecosystem," including dashboards, Prometheus, Loki, alerting, incidents and OnCall. It works with self-hosted Grafana and Grafana Cloud.

Best for: Teams on the Grafana stack who want to keep the server inside their own infrastructure.

Pricing: The server is free under Apache-2.0. Grafana's pricing page lists a Free Cloud tier, Pro "From $19/ month + usage" and Enterprise starting at $25,000 a year.

Pros:

• The --disable-write flag runs it in "read-only mode."
• Read-only mode also removes raw SQL tools that could write.
• Runs through uvx, Docker, a binary or Helm.

Cons:

• You host and secure it yourself.
• Needs a Grafana service account token.

Checked against: Grafana's GitHub repository on 28 October 2026. Full profile: Grafana MCP server.

3. Sentry: production errors and performance

What it is. Sentry's hosted MCP server at https://mcp.sentry.dev/mcp "connects AI assistants to Sentry for searching errors, analyzing performance, triaging issues, reading documentation, and managing projects." You can scope it to one organization or project.

Best for: Developers who want an AI to read a stack trace and its context without copy and paste.

Pricing: No separate MCP price. Sentry's pricing page lists a free Developer plan with one user and 5,000 errors, Team at $29 per month and Business at $89 per month, billed monthly.

Pros:

• "All connections use OAuth."
• Organization and project scoping limits what the AI sees.
• A local package exists for self-hosted Sentry, per our profile.

Cons:

• It can manage projects, not only read them.
• Experimental tools sit behind an opt-in flag.

Checked against: Sentry's MCP page on 28 October 2026. Full profile: Sentry MCP server.

4. PagerDuty: incidents and on-call from chat

What it is. The PagerDuty MCP Server is hosted at https://mcp.pagerduty.com/mcp, with an EU endpoint. Its browse_* tools read and its manage_* tools create, update and delete. "The self-hosted MCP server is now deprecated."

Best for: On-call engineers and incident commanders.

Pricing: No separate MCP price. PagerDuty's pricing page lists a Free plan for up to 5 users and Professional at $25 per user per month, or $21 billed annually.

Pros:

• Hosted in the US and EU.
• Clear split between browse and manage tools.
• Scoped OAuth clients can limit which tools succeed.

Cons:

• Requires an account "with Advanced Permissions."
• PagerDuty says tool filtering "is not directly available" on the hosted server.

Checked against: PagerDuty docs on 28 October 2026. Full profile: PagerDuty MCP server.

5. New Relic: NRQL and investigations

What it is. New Relic's MCP server lets agents "query, investigate, or act on New Relic data." It is hosted at https://mcp.newrelic.com/mcp/, with EU and Japan endpoints. OAuth-only tools include turning plain English into NRQL and analyzing deployment impact.

Best for: New Relic customers who want natural-language queries and deployment analysis.

Pricing: No separate MCP price. New Relic's pricing page includes "100 GB data ingest" a month and one free full platform user, then $0.40 per GB on the original data option.

Pros:

• Three regional endpoints.
• OAuth with an observability:read scope.
• Supports Claude Code, Codex, Claude Desktop and VS Code, among others.

Cons:

• Some tools are OAuth-only and missing with an API key.
• Some tools are preview and need extra opt-in.

Checked against: New Relic docs on 28 October 2026.

6. Honeycomb: tracing and high-cardinality queries

What it is. Honeycomb's MCP server is hosted at https://mcp.honeycomb.io/mcp, with an EU endpoint. Honeycomb says "Most tools are read-only." Write tools need the mcp:write scope and can create or update Boards, Triggers, SLOs and investigations.

Best for: Teams that debug with distributed tracing and wide events.

Pricing: No separate MCP price. Honeycomb's pricing page lists a Free plan with up to 20 million events a month and Pro "Starting at $150 per month."

Pros:

• Mostly read-only, with writes behind a separate scope.
• Honeycomb says to "Prefer OAuth in most use cases."
• US and EU endpoints.

Cons:

• Your team must enable Honeycomb Intelligence first.
• Only team owners can create the API keys for unattended agents.

Checked against: Honeycomb docs on 28 October 2026.

7. Dynatrace: DQL and problem analysis

What it is. The Dynatrace MCP server is remote, served from each environment's own URL. Dynatrace lists uses such as generating and running DQL queries, investigating problems and vulnerabilities, analyzing Kubernetes events and forecasting time series.

Best for: Enterprises on Dynatrace who want AI help with DQL and root-cause work.

Pricing: No separate MCP price. Dynatrace's pricing page lists Infrastructure Monitoring at $29 per host per month and Full-Stack Monitoring at $58 per month per 8 GiB host, within an annual commitment. Log queries on the pay-per-query option cost "$0.0035 per GiB-scanned."

Pros:

• Runs inside your Dynatrace environment.
• Bearer tokens with defined scopes.
• Covers problems, vulnerabilities and Kubernetes.

Cons:

• No dynamic client registration, so OAuth needs a client secret.
• AI-run log queries can add scan costs.

Checked against: Dynatrace docs on 28 October 2026.

8. Elastic: observability through Agent Builder

What it is. Elastic's MCP server is an endpoint in Kibana at {KIBANA_URL}/api/agent_builder/mcp. It exposes Agent Builder tools such as ES|QL queries and search. It is "Generally available" on Serverless Observability projects and on Elastic Stack 9.3 and later.

Best for: Teams that keep logs and traces in Elasticsearch.

Pricing: No separate MCP price. Elastic's pricing page describes Serverless as "Usage based pricing" and Hosted as "Resource based pricing," without list prices on the main page.

Pros:

• Built into Kibana, with Kibana Spaces support.
• API keys work on Stack and Serverless.
• OAuth 2.1 on Serverless.

Cons:

• OAuth is limited to Serverless projects.
• Our profile notes some tools need higher license tiers.

Checked against: Elastic docs on 28 October 2026. Full profile: Elastic MCP server.

9. Splunk: an app for the Splunk platform

What it is. Splunk says its MCP Server "provides a standardized, secure, and scalable interface to connect AI assistants" with Splunk data. It became generally available at version 1.0.0. New deployments must use the Splunk MCP Server App from Splunkbase.

Best for: Security and IT teams whose logs live in Splunk.

Pricing: Splunk's MCP page lists no price. We did not find a public list price for the Splunk platform to quote.

Pros:

• Generally available.
• Built-in role-based access control.
• Admins can turn tools on or off globally.

Cons:

• The older cloud-hosted endpoint is deprecated.
• Some tools are still in preview or beta.

Checked against: Splunk docs on 28 October 2026.

Guardrails for AI in production monitoring

Observability data is sensitive. Logs can hold customer emails, tokens and IP addresses, and the same servers can also change alerts and incidents. Before connecting an AI:

1. Start read-only. Use Grafana's --disable-write, Datadog's mcp_read role or PagerDuty's browse tools.
2. Use a service account. Datadog and Honeycomb document service account keys for unattended agents.
3. Pick the right region. Use EU endpoints if your data must stay in the EU.
4. Scrub logs. Mask secrets and personal data before they reach a model.
5. Log the agent. Record every tool call. See MCP audit logging and the MCP security checklist.

Next step

If you want an AI assistant that can read across Datadog, PagerDuty and your runbooks without touching production, our MCP development team can build and lock it down. Not sure where your risks are? Start with an MCP readiness audit.

Frequently asked questions