New Relic Preflight で Claude Code と Cursor の挙動を計測するThe author walks through integrating New Relic's open-source Preflight tool…

Developers who rely on AI coding assistants such as Claude Code and Cursor often lack a clear picture of what those tools actually cost and where they slow down. A blog post on Zenn walks through using New Relic's open-source Preflight tool to turn that vague sense into measurable data, documenting both the setup process and the places where the official instructions do not match reality.

The author frames the problem in familiar terms: they use Claude Code and Cursor every day, but their understanding of how much each tool consumes, what it costs, and where bottlenecks occur has been based only on intuition. Preflight, released by New Relic as open-source software, is designed to instrument exactly that gap. The project is hosted under the newrelic-experimental organization on GitHub, and the author tested version v1.14.34 across both Claude Code and Cursor.

Preflight fits into the broader category of observability tooling, a discipline that New Relic has built its business around. Observability traditionally refers to collecting metrics, logs, and traces from running software so engineers can understand system behavior after the fact rather than guessing. Applying that mindset to AI coding agents is a comparatively new idea. Instead of monitoring a web service, Preflight appears to capture signals about how these assistants operate: token consumption, cost, and the points where a workflow stalls.

The connection to the Model Context Protocol (MCP) is central to why this matters for the current tooling landscape. MCP is an open standard for connecting AI models to external tools and data sources, and it has become a common integration layer for editors and agents including Claude Code and Cursor. Tagging this work under MCP and mcp-server">mcp-server suggests Preflight plugs into that ecosystem, allowing it to sit between the coding assistant and the services it calls. That positioning is what makes a single instrumentation approach viable across two different products.

Token usage is the practical concern the article emphasizes. Large language models are billed by tokens, the chunks of text a model reads and generates, so a developer's real spending is tied directly to how much context each request carries and how often the assistant re-reads files or repeats work. Without measurement, it is difficult to know whether a slow or expensive session stems from a large codebase, an inefficient prompt, or the tool's own behavior. Preflight's stated purpose is to make those quantities visible so they can be compared and optimized rather than estimated.

A notable part of the write-up is the author's candor about friction during installation. They report that following the official procedure did not go smoothly and that the documented steps diverged from what actually happened when installing v1.14.34. This kind of first-hand account is useful precisely because experimental projects often move faster than their documentation. Readers considering the same setup should expect to adapt the instructions, and the fact that the repository lives under newrelic-experimental reinforces that the tool is still evolving.

For context, the timing aligns with a wider industry push to bring cost and performance monitoring to AI-assisted development. As teams adopt agentic coding tools, questions about budget control, latency, and reliability have grown, and vendors across the observability space have started extending their platforms toward large language model workloads. New Relic's decision to release Preflight as open source, referenced in a company press announcement dated to late July, is consistent with that trend and lowers the barrier for developers to experiment without committing to a paid tier first.

The practical takeaway is modest but concrete. If you already work in Claude Code or Cursor and want to replace guesswork with numbers, Preflight offers one path to instrument token usage and locate bottlenecks, provided you are prepared to troubleshoot a setup that is not yet fully polished. As with any experimental release, behavior may change between versions, and the author's experience with v1.14.34 is a snapshot rather than a guarantee. Still, the underlying shift it represents, treating AI coding sessions as something to be measured like any other production system, is likely to become more common as these assistants move deeper into everyday workflows.