A hosted FinOps MCP server for your cloud and AI cost data

What the tools cover

65 tools — 41 read, 24 write — grouped the way the CloudQuell dashboard is, returning the same numbers it shows.

Cost and forecasts

query_costs for month-to-date versus last month by account and service; compare_periods to rank what changed; get_forecast for projected month-end spend with a confidence note; run_cost_pivot to group by region, usage type, tag, cost center, or resource.

Anomalies

list_anomalies and investigate_anomaly take a spike down to the resources behind it, with get_resource_breakdown and get_weekly_cost_slices for the detail. Dismissing an anomaly is a write.

Savings and commitments

get_savings returns the ranked AWS recommendations with banked, open, and potential totals; list_commitments and get_expiring_commitments read utilization and upcoming expiries.

Budgets, alerts, and schedules

Read budget status against forecast, alert rules, firing history, and report schedules — and, with writes enabled, create or edit budgets, alert rules, scheduled report emails, and notification channels.

AI spend

get_ai_spend returns OpenAI and Anthropic spend by model, token type, workspace, or project. Bedrock, Azure AI Foundry, and Gemini / Vertex AI are added only on request, because those dollars are already inside your cloud totals.

Trust and context

get_data_freshness returns a fresh, stale, or degraded verdict per connection before a number reaches a person; get_finops_briefing runs up to eight read tools in one call for a Monday-morning review.

What a question turns into

You ask in plain language; the agent picks the tools and chains them. Most clients let you expand a call to see its inputs and raw output.

How it connects

One URL and one sign-in. CloudQuell resolves your organization and role from your login on every request.

Endpoint

https://mcp.cloudquell.com/mcp — a remote, hosted server over stateless Streamable HTTP, running on AWS Lambda and API Gateway in us-west-2. Nothing to install or run.

Clients

Claude on the web and Desktop, Claude Code, and ChatGPT are tested end to end. Cursor and Codex connect through the mcp-remote bridge, Goose as a remote extension, and any client with Streamable HTTP and OAuth works.

Sign-in

OAuth 2.1 with PKCE through the CloudQuell sign-in page, by email and password or Google. Discovery documents (RFC 9728, RFC 8414) and a Dynamic Client Registration shim mean a compliant client needs no pre-issued client id. Access tokens last about an hour.

Headless keys

For scripts, CI, and scheduled agents, each member of a paid organization can create one cak_ API key under Profile → API access. Keys are read-only by design — even an admin's — valid for one year, stored only as a hash, and revocable at any time.

Writes

Off by default for every organization. (The one exception, switch_organization, only changes which of your organizations the agent works in and needs just confirm: true.) Any other write needs all four: the admin:write scope, MCP writes enabled by an org admin under Settings → Organization → AI agent access, an admin role, and confirm: true on the call — checked server-side. An agent cannot turn writes on over MCP.

Data scope

Only organizations you belong to, one at a time. The agent reads CloudQuell's already-ingested cost data; it never sees AWS keys, role ARNs, or connection secrets, and the server never touches your cloud accounts. Cost data is ingested daily, and provider billing exports lag two to three days.

Versus AWS's server

AWS's open-source Billing and Cost Management MCP Server runs on your machine through uvx or Docker, signs every call with your own AWS credentials, and covers AWS only; AWS notes that each API call it makes is charged to your account. CloudQuell's server is hosted, signs in with OAuth, and answers from six providers' data already in one ledger.

Plans

MCP tool calls are on every paid plan: Starter $99/month up to $50K of monthly spend, Growth $199/month up to $200K. A Free organization can connect and list the tools, but every cost and admin tool call returns upgrade_required.

What it unlocks

Tools answer one question at a time. These make the routine repeatable and the answers checkable.

• Six FinOps skills

  Weekly briefing, spend review, anomaly triage, budget guardrails, savings actions, and allocation hygiene — runbooks in the open Agent Skills format that tell the agent which tools to call, what to double-check, and when to stop and ask. The server also serves all six as MCP prompts.

  FinOps skills (docs) →

• An agent for cost optimization

  Ask for the top savings opportunities above a dollar floor, cross-check purchase recommendations against what is expiring, and — once writes are on — update a recommendation status with your confirmation. Savings recommendations cover AWS, and no tool changes a cloud resource.

  Cloud cost optimization →

• Answers you can check

  An oversized result comes back as truncated: true with instructions to narrow the request, never as silently partial data, and anomaly signals are labeled as derived from period-over-period movement. Have the agent restate a write before it runs, then re-read to confirm it landed.

  Validating responses (docs) →

• AI spend without double counting

  get_ai_spend keeps direct OpenAI and Anthropic spend separate from cloud-billed AI — Bedrock, Azure AI Foundry, Gemini / Vertex AI — which already sits inside your cloud totals, so an agent's AI figure is never added on top of the cloud bill.

  AI cost management →

Common questions

What is a FinOps MCP server?

A Model Context Protocol server that exposes cost data as tools an AI client can call — query costs, compare periods, list anomalies, read budgets and savings recommendations. You ask "why did our AWS bill jump last week?" in Claude, ChatGPT, or Cursor; the agent chooses the tools, chains the calls, and answers from your billing data rather than its training set, with the numbers it used.

Is there an AWS Billing and Cost Management MCP server?

Yes. AWS publishes an open-source Billing and Cost Management MCP Server in its awslabs/mcp repository. It runs on your machine through uvx or Docker, uses your own AWS credentials, and covers Cost Explorer, Budgets, Cost Anomaly Detection, Cost Optimization Hub, Compute Optimizer, AWS Pricing, S3 Storage Lens, and more — AWS only. AWS states that each API call it makes is charged to your account; Cost Explorer requests are $0.01 each. CloudQuell's server is hosted with OAuth sign-in and covers AWS, Azure, Google Cloud, Snowflake, OpenAI, and Anthropic in one normalized ledger, with budgets, anomalies, and allocation on top. For AWS-only teams comfortable managing credentials locally, AWS's server reaches AWS APIs CloudQuell does not expose, such as pricing and Storage Lens.

Can an AI agent change anything through the CloudQuell MCP server?

Not until your organization allows it. MCP writes are off by default for every organization. A write — creating a budget or alert rule, editing a notification channel, sending an invite, dismissing an anomaly, updating a recommendation status — needs the admin:write scope, writes enabled by an org admin, an admin role, and confirm: true, all checked server-side. Writes change objects inside CloudQuell only; no tool modifies your cloud resources. API keys are read-only.

Which AI clients work with the CloudQuell MCP server?

Claude on the web and Desktop, Claude Code, and ChatGPT are tested end to end. Cursor and Codex connect through the mcp-remote bridge, Goose as a remote extension, and any client that supports Streamable HTTP with OAuth 2.1 and PKCE configures itself from the server's discovery documents. Setup takes about five minutes.

Can I run a scheduled or headless FinOps agent?

Yes. A cak_ API key sent as a bearer token needs no browser sign-in and is read-only, so an unattended job — a daily get_finops_briefing run, say — cannot change anything even if the key leaks. Each member holds one key at a time, valid for one year and revocable by its owner or an org admin. For an automation that needs writes, use an OAuth connection signed in as you, such as a claude.ai scheduled agent.

Is the CloudQuell MCP server free?

No — MCP tool calls are a paid feature. A Free organization can connect and list the 65 tools, but every cost and admin tool call returns upgrade_required. Every paid plan includes MCP: Starter $99/month for up to $50K of monthly cloud spend, Growth $199/month up to $200K, flat and never a percentage of your bill. Both start with a 14-day trial and no credit card.