FinOps MCP servers in 2026: CloudQuell, Vantage, AWS, Finout and CloudZero compared

Cost tools were built around dashboards: a person opens a view, sets a date range, groups by service, and reads the chart. That works when the question fits the view. Most real questions do not. "Why did our AWS bill jump last week?" means checking the period comparison, then the services that moved, then the anomalies, then the accounts behind them — four views and a mental join.

The Model Context Protocol (MCP) changes who does that legwork. It is an open standard for connecting AI assistants to external tools, and in the last year most FinOps vendors, and AWS itself, have published an MCP server for cost data. With one connected, you ask the question in Claude or Cursor; the assistant calls the period-comparison, anomaly and breakdown tools itself, chains the results, and returns the explanation — which service, which account, one-off or trend — in the same chat, with the numbers it used.

The servers are not interchangeable. They differ on where they run, whose credentials the assistant holds, whether it can change anything, and which bills it can see. This post explains those differences, then compares the servers you can connect today, with every competitor claim taken from that vendor's own documentation (sources are linked at the end). We make one of them, CloudQuell; the table and the fit guide show where the trade-offs differ.

What is a FinOps MCP server?

An MCP server exposes cost data as a set of tools — query costs, list anomalies, read budgets, fetch savings recommendations — that an MCP client such as Claude, Claude Code or Cursor can call while you talk to it. You ask "why did our AWS bill jump last week?"; the agent picks the tools, chains the calls, and answers from your data instead of from its training set. A FinOps MCP server is one whose tools cover billing, allocation and optimization rather than infrastructure operations.

What actually matters when choosing one?

Five properties decide whether a FinOps MCP server fits your team: where it runs, how the assistant authenticates, whether it can write, which providers it covers, and what it costs.

• ◦Hosting — local (a process on each user's machine, over stdio) or hosted (a URL the client connects to).
• ◦Authentication — whether the agent signs in with OAuth or holds a long-lived AWS key or API token in a config file.
• ◦Write safety — whether the agent can change budgets, alerts or recommendations, and what has to be true before it does.
• ◦Coverage — AWS only, or cloud, data-warehouse and AI provider spend in one server.
• ◦Price — what it costs to run and whether the price is published.

FinOps MCP servers compared

Six options, starting with ours. Each section covers the five properties above and the setup it suits.

1. CloudQuell — hosted, cloud and AI spend in one server

CloudQuell's server is hosted at https://mcp.cloudquell.com/mcp. You connect with OAuth 2.1 and PKCE through the CloudQuell sign-in page; nothing is installed and no cloud credential reaches the client. It exposes 64 tools — 41 read, 23 write — across costs, forecasts, anomalies, budgets, commitments, savings, tag coverage, allocation, reports and alerts, over AWS, Microsoft Azure, Google Cloud, Snowflake, OpenAI and Anthropic spend in one ledger. It also ships six installable Agent Skills, such as anomaly triage and a weekly briefing, that teach the assistant the full workflow.

Writes are off by default for every organization. A write happens only when the token carries the admin:write scope, an org admin has enabled agent writes, the signed-in user is an admin, and the call includes confirm: true — all checked server-side. Service-account keys for headless agents are read-only.

Pricing is flat and published, never a percentage of spend. CloudQuell is free under $10K/month of tracked spend, without MCP. MCP is on paid plans: $99/month for tracked spend under $50K, and $199/month from $50K to $200K.

CloudQuell fits teams that want cloud, Snowflake and AI provider spend behind one hosted OAuth server at a published price. Its limits: six providers is fewer than Vantage or Finout connect, so Databricks, Datadog or Kubernetes cost is out of scope; savings recommendations cover AWS only for now; and the Free plan does not include MCP tool calls.

2. Vantage MCP — the widest provider list

Vantage runs a hosted server at https://mcp.vantage.sh/mcp over Streamable HTTP with OAuth 2.1, and accepts a Vantage API token as a bearer token as a fallback. A self-hosted version runs locally over stdio with npx -y vantage-mcp-server and an API token. The tools read and write: the open-source repository includes create, update and delete tools for cost reports and budgets, plus anomalies, forecasts, recommendations, virtual tags and dashboards.

Vantage's integration list is the broadest here — AWS, Azure, Google Cloud, Kubernetes, Datadog, Snowflake, Databricks, OpenAI, Anthropic and more. Its pricing is tiered by tracked spend: free under $2,500/month, Pro at $30/month up to $7,500, Business at $200/month up to $20,000, and a custom Enterprise quote above that. Vantage lists the MCP server as a plan feature; check its pricing page for which tiers include it.

Vantage makes more sense when provider breadth is the deciding factor. The trade-off is price at scale: a team tracking $40K a month is past the Business tier and into a sales quote; at that spend level CloudQuell's published price is $99/month.

3. AWS Billing and Cost Management MCP Server — official, AWS only, local

AWS publishes an open-source server in the awslabs/mcp repository. It runs on your machine — installed with uvx or Docker and connected over stdio — and every call uses your own AWS credentials, from aws configure, environment variables or a named AWS_PROFILE. It covers Cost Explorer, Budgets, Cost Anomaly Detection, Cost Optimization Hub, Compute Optimizer, the Pricing API, S3 Storage Lens, Billing Conductor, cost allocation tags, cost categories and more, and can analyze costs across linked accounts. It replaces the older Cost Explorer MCP server, which AWS has deprecated.

The server is free; the calls are not. AWS states that each API call the server makes is charged to your account, and Cost Explorer bills $0.01 per request — an agent chaining several queries per question adds up on a busy day. AWS does not describe the server as read-only: its documented IAM permissions include creating Athena workgroups and tables for Storage Lens queries.

It fits teams that run only AWS and whose users are engineers comfortable managing IAM credentials locally. It does not see Azure, Google Cloud, Snowflake or AI provider spend, and every user needs AWS billing permissions on their own machine. AWS also launched a managed, remote AWS MCP Server in May 2026; it is a general-purpose server that can call AWS APIs, not a billing server with cost-specific tools.

4. Finout MCP — hosted and read-only

Finout's server is hosted only, at https://mcp.finout.io/mcp — there is nothing to install. It supports OAuth for individual users, enterprise AI platforms, machine-to-machine clients and public PKCE clients, and respects the signed-in user's role-based access. Finout states it is currently read-only: it cannot create, edit or delete anything. Its 20-plus tools cover cost queries, anomalies, waste, commitment coverage, tag coverage, budgets and forecasts, and virtual tags.

Finout's MegaBill spans AWS, Google Cloud, Azure, OCI, Snowflake, Databricks, Datadog, OpenAI, Anthropic, Kubernetes and more. Pricing is a flat fee by committed-spend tier, quote only. It makes most sense for existing Finout customers, or where read-only is a hard requirement for every agent.

5. CloudZero MCP — hosted, with an API-key caveat

CloudZero hosts an MCP server that users connect to with OAuth. Its 20-plus tools cover cost queries, dimensions, telemetry, Kubernetes and optimization recommendations. For automation it also accepts an API key, capped at 3,000 tool calls per day — and CloudZero's documentation is explicit that API-key users have full access to all cost data and all tools, with role-based data filters not applied. CloudZero sells on a custom quote with no published rate card. It makes most sense for teams already using CloudZero for unit economics.

Setup runs through CloudZero's AI Hub, and CloudZero also publishes a Claude Code plugin marketplace for its tools. Most tools read; one, which sets the organization context, writes and requires an org-admin permission.

6. Community AWS Cost Explorer servers

Open-source projects such as aarora79/aws-cost-explorer-mcp-server and ravikiranvm/aws-finops-mcp-server predate the official AWS server. Both authenticate with your local AWS credentials and cover AWS only; their last commits were in April and June 2025. They are useful for experimenting; for most teams, the maintained AWS server is the safer default.

If you already run Datadog, Harness, Kubecost or Azure

Several platforms now include cost tools in a broader MCP server. Datadog's MCP server has a cost toolset whose recommendations tool ranks savings by estimated daily impact and requires Cloud Cost Management read access. Harness's official MCP server includes a Cloud Cost Management toolset. Kubecost ships a read-only MCP server in Kubecost 3.3 and later. Microsoft has announced Cost Management and Pricing toolsets for its Azure Resource Manager MCP server. We found no official Google Cloud billing MCP server as of September 2026. If your cost data already lives in one of these platforms, its server is the shortest path.

FinOps MCP servers side by side

Which type fits your setup?

• ◦Cloud, Snowflake and AI provider spend behind one hosted OAuth server, with a published flat price and writes that stay off until an admin opts in — CloudQuell.
• ◦Need Databricks, Datadog or Kubernetes coverage in the same server — Vantage, or Finout for enterprise buyers.
• ◦AWS only, a few engineers, and local IAM credentials are acceptable — the AWS Billing and Cost Management MCP Server.
• ◦Already invested in Finout, CloudZero or Datadog for cost management — that vendor's server; your data and allocation rules are already there.

Connect CloudQuell MCP

• ◦Claude and Claude Desktop: Settings → Connectors → Add custom connector → paste https://mcp.cloudquell.com/mcp.
• ◦Claude Code: run the command below.
• ◦Cursor: add the server URL to mcp.json, as below.
• ◦stdio-only clients such as Codex: bridge with npx -y mcp-remote https://mcp.cloudquell.com/mcp.

claude mcp add --transport http cloudquell https://mcp.cloudquell.com/mcp

// Cursor mcp.json
{ "mcpServers": { "cloudquell": { "url": "https://mcp.cloudquell.com/mcp" } } }

A browser window opens the CloudQuell sign-in. Approve access, then ask: "Using CloudQuell, what is our month-to-date spend versus last month, and what are the top three movers?" CloudQuell is free under $10K/month of tracked spend; MCP tool calls are on paid plans from $99/month.

Is a local or hosted MCP server safer for cost data?

It depends on what the credential can do. A local server needs a cloud credential on every user's machine, and that credential usually reaches more than billing. A hosted server with OAuth keeps the cloud credential with the vendor and gives the client a short-lived token scoped to one server, revocable from one place. The questions to ask either way: can the credential write, and who can revoke it.

Can an AI agent change my budgets or alerts through MCP?

On some servers, yes. Vantage's server has create, update and delete tools; Finout's is read-only. On CloudQuell, writes are disabled for every organization until an admin enables them, and each write also needs an admin role and an explicit confirm: true from the agent.

Does the AWS Billing and Cost Management MCP server cost anything?

The server is free and open source, but AWS charges for the API calls it makes on your behalf. Cost Explorer requests are $0.01 each, and one agent question can trigger several.

Which FinOps MCP servers include OpenAI and Anthropic spend?

CloudQuell, Vantage and Finout all ingest OpenAI and Anthropic cost alongside cloud spend. AWS's server sees model spend only when it is billed through Amazon Bedrock on your AWS bill. CloudQuell's get_ai_spend tool returns OpenAI and Anthropic spend by model, and adds Bedrock only when you ask for it, because Bedrock is already inside your AWS total.