AI coding assistants have become a core part of the modern development workflow. Developers using Claude Code, Cursor, GitHub Copilot, and Windsurf are shipping faster, but for many, that speed can come with a blind spot. Sessions may burn tokens, tokens cost money, and most teams have little visibility into whether that spend is driving results or just generating noise. When a developer ends a long coding session, they typically have no idea how much it cost, whether their AI was working efficiently, or whether the same mistakes are repeating session after session.

New Relic Preflight, initially released as AI Coding Observability is an open-source MCP server that brings full observability to AI-assisted development. It hooks directly into your AI coding assistant and captures structured telemetry, including session metrics, cost breakdowns, behavioral patterns, and efficiency scores. Run it locally with no account required, or connect New Relic to bring that same data into your existing observability platform, queryable with NRQL alongside your production systems.

Let's explore what Preflight makes visible in your workflow.

The hidden cost of AI-assisted development

AI coding assistants are billed by the token. A focused, productive session and an unfocused, meandering one can look identical from the outside, at least until the bill arrives. Preflight tracks token spend at the session, daily, and weekly level, broken down by model and tool type. More importantly, it gives you thresholds to work with. It allows you to configure session, daily, or weekly budget limits and receive alerts before you overshoot. This turns AI spend from an unpredictable line item into something you can plan and manage like any other cloud cost.

Detecting when your AI is working against you

Not all AI activity is productive. Some of the most expensive interactions are also the most wasteful, and they're hard to spot in the middle of a session. The anti-pattern detector identifies four common failure modes:

1. Thrashing: repeatedly calling the same tools in sequence without making progress
2. Re-reads: fetching files the AI already has in context, inflating token use for no reason
3. Blind edits: making code changes without verifying output first
4. Stuck loops: circular decision chains that never resolve

These patterns show up in every developer's workflow occasionally. Seeing them clearly is the first step to breaking them. Over time, identifying which patterns correlate with your most expensive sessions can fundamentally change how you prompt and structure tasks.

A personal efficiency score, automatically

 Preflight generates a weekly coaching report that puts your metrics in context. The efficiency score combines cost per outcome, task completion rate, tool selection quality, and anti-pattern frequency into a single number you can track over time.

The personal coaching report goes even further: it compares your current week against your own historical baseline, surfaces where you're improving, and calls out specific session types that cost the most relative to their outcomes. It's the kind of retro most developers never have time to run manually, and it’s generated automatically at the end of every week.

Works with New Relic, or on its own

Preflight runs fully featured out of the box—no New Relic account required. Local mode gives you the complete dashboard, session history, weekly summaries, efficiency coaching, and alerting, all stored on your machine.

When you're ready to bring that data into your existing observability platform, you can easily connect to your New Relic account. Your AI coding telemetry lands alongside your production systems, queryable with NRQL, visible in cloud dashboards, and managed through New Relic's alerting workflows. For teams, connecting also enables org-wide aggregation via NerdGraph, giving engineering leads a cross-developer view without any per-developer setup.

Getting started in minutes

Preflight installs alongside your existing AI coding assistant. It supports Claude Code, Cursor, GitHub Copilot, Windsurf, Zed, Continue.dev, Amazon Q, and any generic MCP-speaking client.

You can be up and running in under five minutes, this is all it takes to get started:

1. Clone the repository and build: git clone https://github.com/newrelic-experimental/preflight && cd preflight && npm install && npm run build && npm link
2. Run the interactive setup wizard: preflight setup
3. The wizard installs hook scripts, validates your keys against New Relic, and pre-fills your developer name. 
4. Start a session in your AI coding assistant; metrics begin flowing immediately

No New Relic account? Skip the key prompts in the wizard and choose local mode, that way everything stores on your machine and you can connect to New Relic later.