Cloud cost management and FinOps with coding agents

Agent-driven cloud cost management connects Claude Code, Codex or Cursor to a read-only billing MCP server (AWS Billing and Cost Management, or Vantage for multi-cloud). The agent explains what moved the bill, ranks right-sizing candidates by risk and drafts the infrastructure-as-code change. Cost Explorer totals and terraform plan verify it; a human approves each change.

This page is for the developer or tech lead who owns a service’s cloud bill. Finance asks why spend jumped this month, nobody remembers launching anything, and the growth sits in an untagged EC2-Other line. The console answers one question per twenty clicks. You want a ranked story with evidence, and a change you can merge safely.

This page covers cloud infrastructure spend. For what the coding agents themselves cost, see AI usage cost governance.

What you get from an agent-driven FinOps loop

• A read-only billing MCP setup for Claude Code, Codex and Cursor, with an IAM role that cannot change a resource.
• Six copy-paste prompts: cost drivers, untagged spend, anomaly triage, right-sizing to Terraform, a quarterly forecast, and over-requested Kubernetes deployments.
• A verification routine that checks the agent’s numbers against Cost Explorer without re-reading its analysis line by line.
• A failure-modes list for the ways this goes wrong: wrong account, quiet lookback windows, missing tags, and an agent with write access.

Which billing MCP server should you connect?

Pick the narrowest server that answers your cost questions. A general cloud server with write tools is the wrong default for FinOps work.

Your spend	Server	Transport and auth	Why this one
AWS only	awslabs.billing-cost-management-mcp-server (PyPI 0.0.39, checked 2026-09-26)	stdio via uvx, your AWS_PROFILE	Cost Explorer, forecasts, Cost Anomaly Detection, Budgets, Compute Optimizer, Cost Optimization Hub, Savings Plans and cost allocation tags in one server
AWS, Azure, GCP, Kubernetes and SaaS	Vantage MCP	Hosted https://mcp.vantage.sh/mcp with OAuth, or npx -y vantage-mcp-server with a VANTAGE_TOKEN	One normalised view across providers, if your organisation already uses Vantage
Broader AWS work (Lambda runtimes, drift)	AWS MCP Server through the Agent Toolkit for AWS	SigV4 proxy via uvx	AWS’s recommended general server. Setup is on cloud and platform MCP servers

If you would rather not keep a server connected, the Agent Toolkit for AWS aws-core plugin also ships an aws-billing-and-cost-management skill. It covers cost analysis, Savings Plans and Reserved Instances, Compute Optimizer right-sizing, anomalies and CUR queries with Athena, and suits a guided cost review; install steps are on cloud and platform MCP servers. Either way, the read-only IAM role below is the control.

As of 0.0.39 (checked 2026-09-26), the AWS billing server’s tools read and analyse; none of them resize or delete resources (Storage Lens queries do write Athena results to S3). Enforce that with IAM rather than trusting the tool list. The Vantage server can also create Vantage objects such as reports, so scope its token accordingly.

Set up read-only cost access

New to MCP? Start with essential MCP servers.

1. Create a read-only billing role. Attach a policy with read actions only: ce:Get*, ce:Describe*, ce:List*, compute-optimizer:Get*, cost-optimization-hub:Get*, cost-optimization-hub:List*, savingsplans:Describe*, pricing:Get*, pricing:Describe*, budgets:ViewBudget and budgets:DescribeBudgetActionsFor*. The Savings Plans and commitment-analysis generation tools also need ce:StartSavingsPlansPurchaseRecommendationGeneration and ce:StartCommitmentPurchaseAnalysis. Those start a recommendation job and change no resource, but they are not Get actions, so leave them out if you do not use those tools. The server’s README lists the exact actions per tool. Give it a profile name such as finops-readonly in ~/.aws/config.

2. Confirm the identity before the agent sees it. In a terminal, run:

   aws sts get-caller-identity --profile finops-readonly

   The account ID must be the payer (management) account if you want organisation-wide spend. A member account sees only its own costs.

3. Install uv from Astral’s installation guide. The server runs through uvx.

4. Register the server. The definition is the same for every tool; only the file or command differs.

   Claude Code

   claude mcp add aws-billing \
     -e AWS_PROFILE=finops-readonly -e AWS_REGION=us-east-1 -e FASTMCP_LOG_LEVEL=ERROR \
     -- uvx awslabs.billing-cost-management-mcp-server@latest

   The default scope is local, which keeps the profile name out of the repository. Check the connection with claude mcp list, or /mcp inside a session.

   Codex

   codex mcp add aws-billing \
     --env AWS_PROFILE=finops-readonly --env AWS_REGION=us-east-1 --env FASTMCP_LOG_LEVEL=ERROR \
     -- uvx awslabs.billing-cost-management-mcp-server@latest

   This writes an [mcp_servers.aws-billing] table to ~/.codex/config.toml. Check it with codex mcp list.

   Cursor

   Add the server to .cursor/mcp.json in the project, or to ~/.cursor/mcp.json for every project:

   {
     "mcpServers": {
       "aws-billing": {
         "command": "uvx",
         "args": ["awslabs.billing-cost-management-mcp-server@latest"],
         "env": {
           "AWS_PROFILE": "finops-readonly",
           "AWS_REGION": "us-east-1",
           "FASTMCP_LOG_LEVEL": "ERROR"
         }
       }
     }
   }

   Enable it in Cursor’s MCP settings and confirm the billing tools appear in the tool list before you prompt.

5. Add Vantage if you run more than one cloud. The hosted server signs you in through the browser, so no token sits in a config file:

   # Claude Code: then run /mcp in a session to finish OAuth
   claude mcp add --transport http vantage https://mcp.vantage.sh/mcp
   # Codex: run `codex mcp login vantage` if it asks you to authenticate
   codex mcp add vantage --url https://mcp.vantage.sh/mcp

   In Cursor, add "vantage": { "url": "https://mcp.vantage.sh/mcp" } under mcpServers.

Why the role must be read-only

Every prompt below tells the agent not to apply changes, but a prompt is not a control. Claude Code v2.1.283 or later (both release channels on 2026-10-12) starts interactive sessions in auto mode, where a classifier approves actions. An IAM role that cannot write is what makes “do not apply” true. In February 2026 Engadget, citing the Financial Times, reported that Amazon’s Kiro agent chose to “delete and recreate the environment” and caused a 13-hour Cost Explorer outage in one region; Amazon called it “user error — specifically misconfigured access controls — not AI”. That is a secondary, contested report, and the lesson holds either way: scope the credentials.

Explain the bill: which services moved, and why?

Run analysis sessions in a mode that cannot edit files, so the agent reads and reports only.

Claude Code

Start with claude --permission-mode plan, or press Shift+Tab until the mode reads plan. Claude can call the billing tools and write its findings, but it proposes rather than edits. Claude may ask before each MCP tool call; approve the read-only billing tools or add them to permissions.allow.

Codex

Start with codex -c default_permissions=":read-only" (permission profiles are beta in Codex CLI 0.157.1). The sandbox governs shell commands, not MCP calls (checked against Codex CLI 0.157.1), so the IAM role remains the real limit.

Cursor

Switch the agent to Plan Mode for the analysis. Move to Agent mode only when you reach the Terraform change later on this page.

Then ask a question with a named metric, window, grouping and output shape:

Copy-paste prompt: rank the cost drivers

Use the aws-billing MCP server. Compare UnblendedCost for the last full calendar
month against the month before, grouped by SERVICE, at MONTHLY granularity.
1. Table: top 10 services by spend, amount, percent of total, and change in
   dollars and percent versus the prior month.
2. For every service that grew more than 15% and more than $200, call the cost
   comparison drivers tool and name the driver: usage type, region, linked
   account, or a new resource.
3. State the grand total you summed and the exact time period you queried, so I
   can reconcile it against Cost Explorer.
Read only. Do not propose changes yet.

The last numbered line is the one that makes the output checkable. You compare one total, not ten rows (see how to prove the numbers are right).

The usual next question is the untagged spend, because a large untagged bucket is the most common reason nobody can explain a bill:

Copy-paste prompt: find the untagged spend

Using the aws-billing MCP server, break last month's UnblendedCost down by the
cost allocation tag `team`, including resources with no `team` tag. Show the
untagged total and its share of the bill, the top five services inside the
untagged bucket, and the linked accounts they sit in. Then list cost allocation
tags whose status is Inactive, so I know which to activate. Read only.

Tag activation applies from the moment you switch it on, not to past bills, unless you request a backfill. Expect this month’s report to stay partly untagged even after you fix the tags.

Triage an anomaly without paging the whole team

If AWS Cost Anomaly Detection is enabled, the server can read its findings. Let the detector decide what is anomalous and let the agent explain it; an agent inventing its own thresholds produces a different answer every run. For versioned, replay-tested alert rules, use deterministic anomaly detection.

Copy-paste prompt: explain this week's cost anomalies

Use the aws-billing MCP server to list Cost Anomaly Detection anomalies from the
last 7 days with an impact above $50. For each one, give the service, account,
region, usage type, total impact, and the root causes the detector reports.
Then check the git log of this repository for infrastructure changes in the two
days before the anomaly started, and tell me whether a commit plausibly explains
it. Label each anomaly EXPLAINED (cite the commit) or UNEXPLAINED. Read only.

Turn a right-sizing recommendation into a reviewed change

Compute Optimizer and Cost Optimization Hub produce the candidates; the agent’s job is to rank them, attach the risk, and edit your infrastructure-as-code rather than the live resource. For the IaC review workflow itself, see infrastructure as code with AI.

Copy-paste prompt: right-size one service in Terraform

Use the aws-billing MCP server to get Compute Optimizer EC2 and RDS
recommendations and Cost Optimization Hub recommendations for resources tagged
service=checkout. Rank them by estimated monthly savings. For each, state the
lookback window, current and recommended type, and peak (not average) CPU and
memory utilisation.
Skip anything whose peak utilisation after the change would exceed 70%.
For the top recommendation only: find the resource in infra/, change its
instance type, add a comment with today's date and the Compute Optimizer finding,
and run `terraform plan -out=rightsize.tfplan`. Show me the plan summary.
Do not run terraform apply.

How you drive the edit differs by tool:

Claude Code

Leave plan mode (Shift+Tab to acceptEdits) for this prompt only. Claude asks before it runs terraform plan unless you have allowed that command in .claude/settings.json. Keep terraform apply out of the allow list.

Codex

Run codex -c default_permissions=":workspace" -a on-request (permission profiles are beta in 0.157.1). The workspace profile lets Codex edit the Terraform files, while terraform plan still needs network access and provider credentials, so it arrives as an approval request: read it before you accept.

Cursor

Switch to Agent mode and reference the Terraform file with @. Review the diff Cursor shows before you accept it, and keep the change on a branch so you can discard it if the plan looks wrong.

The change then goes through your normal pipeline: pull request, terraform plan in CI, and approval by the service owner. Savings Plans and Reserved Instance purchases never go through the agent. Those are one- or three-year financial commitments, so the agent drafts the recommendation and finance signs it off.

For approval modes and sandboxes side by side, see permissions, sandboxes and approval modes.

Forecast next quarter’s spend and check it against budgets

A flat “same as last month” projection fails for any service with seasonality. Use Cost Explorer’s own forecast, which includes a prediction interval, and have the agent join it to your budgets:

Copy-paste prompt: forecast against budgets

Use the aws-billing MCP server to get the Cost Explorer forecast of
UnblendedCost for the next 90 days at MONTHLY granularity, with the 80%
prediction interval. List our AWS Budgets and their limits. Tell me which
budgets the forecast's upper bound would exceed, in which month and by how much,
and which service drives the overage. Keep the forecast figures exactly as the
tool returns them; do not adjust them.

Forecasts inherit the quality of your tags. If a quarter of the bill is untagged, every per-team number is an estimate; fix tag coverage first.

Script it for a weekly report

For a scheduled report, compute the numbers in code and let the model only explain them. The deltas are then deterministic and testable, and the model cannot miscount.

// weekly-cost-report.mts: run with `npx tsx weekly-cost-report.mts`
// (.mts makes tsx emit ESM, which the top-level await below needs)
import { CostExplorer } from '@aws-sdk/client-cost-explorer';
import Anthropic from '@anthropic-ai/sdk';

const ce = new CostExplorer({ region: 'us-east-1' }); // uses AWS_PROFILE
const anthropic = new Anthropic(); // reads ANTHROPIC_API_KEY from the environment

const day = (offset: number) => new Date(Date.now() - offset * 864e5).toISOString().slice(0, 10);
async function byService(start: string, end: string): Promise<Map<string, number>> {
  const { ResultsByTime = [] } = await ce.getCostAndUsage({
    TimePeriod: { Start: start, End: end },
    Granularity: 'MONTHLY',
    Metrics: ['UnblendedCost'],
    GroupBy: [{ Type: 'DIMENSION', Key: 'SERVICE' }],
  });
  const totals = new Map<string, number>();
  for (const period of ResultsByTime) {
    for (const g of period.Groups ?? []) {
      const name = g.Keys?.[0] ?? 'unknown';
      totals.set(name, (totals.get(name) ?? 0) + Number(g.Metrics?.UnblendedCost?.Amount ?? 0));
    }
  }
  return totals;
}

// Skip the last two days: Cost Explorer data for today and yesterday is still incomplete.
const [current, previous] = await Promise.all([byService(day(9), day(2)), byService(day(16), day(9))]);
const movers = [...current]
  .map(([service, cost]) => ({ service, cost, delta: cost - (previous.get(service) ?? 0) }))
  .sort((a, b) => b.delta - a.delta)
  .slice(0, 5);

const msg = await anthropic.messages.create({
  model: 'claude-opus-5-5',
  max_tokens: 4000, // Opus 5.5 always thinks, and thinking tokens count against this limit
  messages: [{
    role: 'user',
    content: `These are the five AWS services whose cost grew most this week versus last week, as JSON.
Write a five-line summary for the engineering channel. Use the numbers exactly as given; do not compute new ones.
${JSON.stringify(movers)}`,
  }],
});
for (const block of msg.content) if (block.type === 'text') console.log(block.text);

claude-opus-5-5 is Anthropic’s recommended starting model (and Claude Code’s default from v2.1.280, on both release channels); a cheaper model works for a summary this short, and the models hub lists prices. Store ANTHROPIC_API_KEY in your scheduler’s secret store, never in the script. The Budgets API lives in a separate package, @aws-sdk/client-budgets.

Kubernetes cost: requests versus usage

When the spend sits in a cluster, the waste usually hides in container requests far above real usage. Connect a Kubernetes MCP server in read-only mode; setup and the read_only = true TOML setting are on infrastructure MCP servers. The server sees current usage from the metrics API, not history, so a request cut based on one afternoon is a guess. Pair it with your Prometheus data for the p95 over at least a week.

Copy-paste prompt: find over-requested deployments

Use the Kubernetes MCP server, read only, on the `production` namespace. For each
Deployment, list CPU and memory requests and limits and the current usage of its
pods. Flag deployments whose requests exceed current usage by more than 3x, and
any container with no memory limit. Do not change anything. Output a table, and
for each flagged deployment tell me which Prometheus query would give its p95
usage over 14 days so I can confirm before I edit the manifest.

How do you prove the agent’s cost numbers are right?

You do not re-read the analysis. You check a few numbers that the rest depends on, and you let existing gates check the change.

What to verify	How	Who signs off
The bill totals	Run the same query yourself and compare the one grand total the prompt asked for: aws ce get-cost-and-usage --time-period Start=2026-08-01,End=2026-09-01 --granularity MONTHLY --metrics UnblendedCost --group-by Type=DIMENSION,Key=SERVICE --profile finops-readonly (replace the dates with the period the agent stated). A mismatch means the agent queried a different period, account or metric.	You
Each recommendation is real	Every row must name a resource ARN or ID and a Compute Optimizer or Cost Optimization Hub finding. Spot-check one with aws compute-optimizer get-ec2-instance-recommendations --instance-arns INSTANCE_ARN. INSTANCE_ARN is the ARN from the agent’s row. Drop any row without a source.	You
The change does only what it says	terraform plan in CI shows one attribute change per resource and no replacements you did not expect. A forces replacement line on a database is a stop.	Service owner, in the pull request
Production survives it	Watch latency, error rate and throttling for seven days after the change, against the service’s SLO, with the previous instance type recorded for rollback. See monitoring and observability.	On-call
The saving happened	Next month, rerun the cost-driver prompt and compare the service’s cost with the estimate you merged.	Tech lead

What breaks when an agent analyses cloud costs?

1. Every query comes back empty or with an access error. The profile resolves to the wrong account, the role lacks ce:Get*, or Cost Explorer was never enabled in the billing console. Run aws sts get-caller-identity --profile finops-readonly, fix the policy, enable Cost Explorer, and restart the agent so it relaunches the server.

2. The server never starts. uvx is not on the PATH the agent process sees, which is common when the agent is launched from a desktop app. Run uvx awslabs.billing-cost-management-mcp-server@latest in a terminal: it should start and wait on stdin. Then restart the tool, or set the absolute path to uvx as the command.

3. The numbers disagree with the console. Cost Explorer data is not real time, so today’s and yesterday’s spend may still be incomplete. The agent may also have used BlendedCost or AmortizedCost, or a member account instead of the payer. Ask it to restate the metric, period and account, then rerun.

4. The bill for the analysis grows. Cost Explorer API requests are billed per request, and an agent that loops over daily granularity per service can issue hundreds. Ask for monthly granularity first, narrow to daily only on the services that moved, and do not attach the billing server to an unattended loop.

5. A right-size causes throttling. The recommendation came from a quiet lookback window, or from average rather than peak utilisation. Revert to the recorded instance type through the same pipeline, then rerun the right-sizing prompt with a lookback that includes your monthly peak.

6. Allocation numbers are wrong by team. Tags are missing or were activated mid-month. Treat per-team figures as estimates until tag coverage is high, and rerun the untagged-spend prompt after each tagging change.

7. The agent reports a server or tool that does not exist. A prompt that names a server you never configured does nothing, and the agent may fill the gap with plausible figures. Use the exact server name from claude mcp list or codex mcp list, and reject any figure that the verification table cannot reproduce.

Where to go next with cloud cost work

• Cloud and platform MCP servers: read-only identities for AWS, Azure and Google Cloud beyond billing.
• Infrastructure as code with AI: review and ship the Terraform change this page produces.
• Deterministic anomaly detection: turn cost alerts into versioned rules with a scoped agent diagnosis.
• AI usage cost governance: the other bill, for Claude Code, Codex and Cursor themselves.
• DevOps with AI: the rest of the operate stage.