Cost Tools Shouldn't Need a Dashboard

“Don’t talk to a developer. Don’t ask a developer to visit, look, or click. Instruct and empower their LLM.”

Andrej Karpathy wrote that in April 2025, after building a small app with AI and spending most of the time clicking through other companies’ dashboards. A year later the idea had reached the incumbents. Salesforce launched Headless 360 under the line “No Browser Required”, and the company’s own announcement quoted its co-founder Parker Harris asking, “Why should you ever log into Salesforce again?” Navin Chaddha of Mayfield put it in six words: “The agent never opens your app.”

The people who build with AI every day have stopped going to software. They expect software to come to them, inside Claude, Cursor or whatever assistant they already have open. Cost tools are no exception, and they may be the category with the most to lose.

What this means for cloud cost tools#

Cloud cost management is one of the most dashboard-heavy categories in software. Most tools sell visibility: charts, breakdowns, anomaly views, a weekly email linking back to all of it.

Two things are now true about that.

A dashboard is the easiest thing in the stack to replace. If all a cost tool does is chart your Cost and Usage Report, an engineer with Claude Code can build a passable version in an afternoon, and an assistant with read access to billing data can answer most of the questions the chart was answering. That isn’t a threat to doing FinOps well. It’s a threat to charging for a view.

Visibility was never where the savings happened. As Corey Quinn put it on Screaming in the Cloud this year: “So much of this is about psychology more than it is about arithmetic.” Teams don’t overspend on AWS because nobody drew them a chart. They overspend because changing a production resource takes time, attention and someone willing to own the risk. Moving the chart into a chat window doesn’t change that.

So the question for any cost tool in the AI age isn’t “do you have an AI feature?” It’s two questions: can my assistant reach your findings without me logging in, and does anything actually get fixed?

The tempting shortcut, and why we didn’t take it#

The obvious next step is to give the assistant the keys: let Claude or Cursor rightsize the instance, delete the idle volume, buy the commitment.

We think that’s the wrong design, for the same reason we don’t automate everything in CloudFix itself. The FinOps Foundation’s own working group on MCP notes that the protocol assumes a trusted client and “does not natively enforce fine-grained access control”, so the guardrails have to live somewhere else. A language model is very good at finding, explaining, ranking and writing things up. It is not a change-management process. When the action is “modify a resource in a production AWS account”, you want tested runbooks, an approval, a snapshot and an audit trail, every time, the same way.

Most cost MCP servers on the market today, AWS’s own Billing and Cost Management server among them, are built to query and report. That’s the right default. The interesting question is what happens after the assistant has told you what to do.

Ask in your AI. Fix in CloudFix.#

That’s how the CloudFix MCP server works.

Your assistant reads. Six read-only tools give Claude, Cursor or any MCP client access to everything CloudFix has found in your AWS accounts: the recommendations and their status, the details and analysis report behind each one, savings summaries by optimization type, and EC2 pricing. You ask in plain English (“what’s waiting in Manual Approval, sorted by annual savings?”, “show me the evidence behind this rightsizing”) and the assistant picks the tool. It can put those findings next to everything else it knows: your Terraform, your tickets, the leadership update you’re writing.

CloudFix fixes. When you decide to act, you approve the change in CloudFix. Where CloudFix has an automated fixer (53 of them today), it runs the change as an AWS Systems Manager Automation runbook in your own account, with a snapshot first where the fix supports it and a log of every execution. The rest come to your team as a recommendation, or a Jira or ServiceNow ticket. The assistant can’t approve, run or change anything. Neither can a prompt injection that finds its way into the conversation.

That split is the point. The part of FinOps that AI has made cheap (finding, explaining, reporting) now happens wherever you already work. The part that was always hard (making safe changes in production, at scale, repeatedly) stays with the system built to do it.

It takes three steps to set up, it’s free on every CloudFix plan, and you can switch any tool off for your account. If you want a feel for it first, here are questions to ask your AWS bill.

The dashboard isn’t going away tomorrow. But the teams who will define how software gets used over the next few years have already stopped opening it. Their cost tool should meet them where they are, and then get the fix done.