New Relic MCP Server

Connect your AI assistant to New Relic — safely.

A production-oriented, self-hostable Model Context Protocol (MCP) server that gives Claude, Cursor, Codex, and other MCP clients typed, read-only-by-default access to NRQL, logs, metrics, traces, errors, entities, alerts, incidents, dashboards, synthetics, workloads, service levels, and selected administration APIs — all over New Relic's NerdGraph.

It deliberately does not expose arbitrary GraphQL, forward client credentials to New Relic, or accept New Relic keys as tool arguments. Mutations are absent unless their feature gates are explicitly enabled, and every write uses a dry-run, confirmation phrase, and post-write readback workflow.

Highlights

• 🔒 Safe by default — read-only out of the box; writes are opt-in, gated, and run a dry-run → confirmation-phrase → post-write readback flow.
• 🧱 Bounded API surface — a fixed, documented set of NerdGraph operations; no arbitrary GraphQL and no credentials passed as tool arguments.
• 🔌 Works with any MCP client — Claude Code, Claude Desktop, Cursor, Codex, and more, over stdio or stateless Streamable HTTP.
• 📈 Usage telemetry built in — a Prometheus /metrics endpoint with per-tool call counts, durations, and outcomes, plus upstream NerdGraph, queue-depth, retry, rate-limit, and cache metrics.
• 🏢 Enterprise-ready — bearer/OIDC authentication, Host/Origin allowlists, account allowlists, secret and PII redaction, Docker, and Helm.
• 🧰 Broad coverage — independently selectable toolsets for NRQL, entities, alerts, dashboards, synthetics, workloads, service levels, logs, and metrics.
• 📚 Documented and tested — architecture docs, ADRs, and a unit/integration/contract/security/load test suite that gates every change in CI.

What you can ask

Once connected, ask your MCP client in plain language — it selects and calls the right read-only tools for you:

• "What are the slowest transactions for the checkout service in the last hour?"
• "Summarize the open incidents for the payments team and their likely causes."
• "Which service levels are at risk of missing their objective this week?"
• "Review my alert policies and point out coverage gaps."
• "Why did the login synthetic monitor fail overnight?"

See the tool catalog for the complete set of capabilities.

Why this MCP?

New Relic offers an official, New Relic-hosted MCP server (New Relic AI MCP, currently in preview). This project is a complementary, self-hostable and open-source alternative. Choose it when you want:

• Self-hosting and control — run the MCP bridge inside your own network with your own credentials; only outbound NerdGraph calls leave your perimeter. (New Relic's preview is not offered for FedRAMP-regulated accounts.)
• Read-only by default, with gated writes — every mutation is opt-in and runs a dry-run → confirmation-phrase → readback flow.
• A fixed, auditable operation surface — a documented set of NerdGraph operations, with no arbitrary GraphQL and no New Relic keys passed as tool arguments.
• Your own controls — bearer/OIDC auth, Host/Origin and account allowlists, secret redaction, and independently selectable toolsets.
• First-class telemetry — Prometheus metrics for tool usage and upstream calls that you can scrape into your existing observability stack.
• Apache-2.0 and forkable — inspect, pin, and extend exactly what you deploy.

If you

... View full README on GitHub