The Infracost MCP server gives any AI assistant, team chatbot or autonomous agent your FinOps and tagging policies, guardrail history, policy violations in code and investigated cloud waste, so it can check, explain and fix your cloud costs.
Not an Infracost customer yet? Test out the MCP with a free 14-day trial of Infracost Cloud.
Your engineers write infrastructure with Claude and Codex. Your platform team runs agents that work through the backlog overnight. Someone asks the team chatbot why the bill jumped. Every one of them is making cost decisions without the context your FinOps practice has built up.
They don't know your tagging and FinOps policies, or which cost increases your team already approved. They might see what a resource costs, but not where it's declared in your code or which pull request put it there. So they guess. Every wrong guess costs you a review cycle, a rewrite, or a line on the bill nobody planned for. You spend more and fix less.
Our new MCP server ends the guessing. It makes Infracost the FinOps context engine for your agents, at every phase of the infrastructure lifecycle: while code is written, while it's reviewed, and after it's running in your cloud. Your rules, what's broken, and how to fix it, available in any AI tool or agent your team uses.
What the Infracost MCP server gives your AI tools and agents
When you connect Infracost's MCP server, your AI tool or agent can reach four things:
Your rules: FinOps, tagging, and security policies, plus the cost guardrails that hold a pull request for approval.
What's broken in your code: policy violations Infracost finds by scanning your repos.
What's wasting money in your cloud: findings our agents have already investigated, each with evidence, the pull request that caused it, and a proposed fix.
Guardrail history: every cost increase that hit a guardrail, and who approved it. You and your agents don't waste time investigating spikes that can already be explained, and you can trace any change back to the person who signed off.
Then it can act. It drafts fixes as pull requests or tickets, and records the outcome.
That context compounds. Every policy you configure and every guardrail approval your team makes becomes something every AI tool and agent your team uses can draw on. The longer Infracost runs, the more they know.
Write Terraform that follows your FinOps and tagging policies
For engineers writing infrastructure with an AI assistant
You're adding a new RDS instance. Your org has tagging policies, FinOps rules and a guardrail that blocks big cost increases, and you don't know them by heart. You don't need to. Ask first:
I'm about to write Terraform for a new RDS instance in us-east-1. What Infracost policies and guardrails apply, and what values should I use?
Your assistant pulls the tags you can't merge without, the FinOps policies for RDS, and the guardrail that could block your merge. It writes Terraform that follows them, then asks for the one thing it can't infer: which service owns the database.
The pull request passes policy on the first commit, not after a round of review comments.
Find the root cause of a cloud cost spike, and whether it was approved
For FinOps teams
The bill jumps. Before anyone escalates, ask:
Are there any unexpected cost spikes on my cloud bill right now? For the biggest one, explain what caused it and whether that change was already approved through a guardrail.
In this example, the answer traces a $2,310/mo spike to a single pull request that pointed a bucket's access logs back into itself, so every logged write generated another. It names the file, proposes a one-line fix, and confirms the change never went through a guardrail approval.
That guardrail check is what saves you the time. Guardrail history is an audit trail of every cost increase your team reviewed before it shipped, and who approved it. If a spike was planned and approved, you know in seconds and skip the investigation. If it wasn't, you hand engineering a cause and a fix instead of a question.
You don't need a terminal or a dashboard for any of this. It works in Claude, ChatGPT, or the chatbot your team already uses in Slack or Teams.
Automate cloud cost fixes with your platform agent
For platform teams running their own agents
Your platform agent already rotates certificates and patches dependencies. Connect it to Infracost, and cost work joins the same queue. Give it one standing instruction:
Every day, check Infracost findings. For each with potential savings over $300 a year, open a fix PR and notify the resource owner.
Each finding arrives already investigated, with the monthly savings, the evidence, the file to change and the effort involved. Your agent spends its run fixing, not rebuilding that analysis from raw billing data.
You stay in control. Fixes are drafted as pull requests or tickets, never as direct changes to your cloud. Agents act with the role of whoever connected them, so you decide whether they only investigate or also draft fixes. You review the pull requests, which you were already doing.
Enforce cost guardrails on every pull request
The MCP server gives your AI tools and agents the rules. Infracost CI enforces them.
Both run on the same policies and guardrails, so every pull request is checked against exactly what your AI tool or agent worked from, whoever or whatever wrote the code. CI prices each change. Under the guardrail, it merges. Over it, the change waits for a person to approve.
Agents move fast. With CI on every pull request, no cost increase or policy violation slips through unnoticed.
How to connect the Infracost MCP server
The MCP server is available to every Infracost Cloud customer today. It's hosted, so there's nothing to install or update, and it runs anywhere: AI assistants, shared chatbots, and agents in sandboxed environments.
Claude: your Claude org owner adds Infracost as a custom connector, then everyone signs in with their Infracost account.
Claude Code, ChatGPT and other MCP clients: point them at
https://mcp.infracost.ioand sign in.Agents and bots: create a service account token in Infracost Cloud.
Frequently Asked Questions (FAQs)
Which AI tools work with the Infracost MCP server?
The Infracost MCP server works with Claude, Claude Code, ChatGPT and any other AI tool or agent that supports MCP. See the setup guide for how to connect yours.
Do I need to install anything to use the Infracost MCP server?
No. The Infracost MCP server is hosted by Infracost, so there's nothing to install or update. It works in AI assistants, shared chatbots, and agents running in sandboxed environments that can't install software.
Can an agent change my cloud through the Infracost MCP server?
No. Agents connected to the Infracost MCP server never change your cloud directly. If you allow them to, they can draft fixes as pull requests or tickets, and a person reviews them before anything changes.
What can an agent do with the Infracost MCP server?
An agent connected to the Infracost MCP server acts with the Infracost role of whoever connected it. For example, a Viewer's agent can read policies, guardrails, issues and findings, and an Editor's agent can also draft fixes and record the outcome.
How is the Infracost MCP server different from the Infracost CLI?
The Infracost CLI runs locally on your machine. The Infracost MCP server is hosted, so any AI tool or agent can reach it with nothing to install, including team chatbots and sandboxed agents. It gives them context such as your policies, guardrail history, issues and investigated findings, and lets them draft fixes.
Is the Infracost MCP server included in Infracost Cloud?
Yes. The Infracost MCP server is available to every Infracost Cloud customer at no extra cost.
Infracost ROI Report
Learn how the ROI of shifting FinOps left is measured