An AI assistant answers questions about cloud cost by calling the provider's cost APIs through an integration, often through the Model Context Protocol (MCP), and writing the result into its reply. The reply is therefore only as correct, current and complete as the query behind it: which scope, which period, which cost measure, and with whose permissions. AWS, Microsoft and Google offer assistants in their consoles for this, and AWS and Google also publish MCP servers that reach cost data. The same test applies to each of them: can every figure in the answer be traced to a call?
How it works
MCP is an open standard through which an AI application uses the data and functions of other systems. The specification, currently in its revision of 2026-07-28, has three roles: the host (the AI application), the client (one connection inside the host) and the server (the service that offers data and functions). Messages follow JSON-RPC 2.0, and the revision describes the protocol as stateless: every request carries everything the server needs to handle it instead of relying on a session. A server provides resources, prompt templates and tools the model can call. For connections over HTTP the specification sets out its own authorization scheme; a server started locally over stdio takes its credentials from the environment. Among its security principles, no tool runs without the user's explicit consent, and a host trusts a tool's description only as far as it trusts that tool's server. Since 2025-12-09 MCP has sat with the Linux Foundation's Agentic AI Foundation, to which Anthropic contributed the protocol.
An answer is built in five steps:
- The user asks a question, for example last month's cost by service.
- The model picks a tool and sets its parameters: scope, period, grouping, cost measure.
- The tool calls the provider's API with the credentials of the user or of the server.
- The result enters the model's context as tokens.
- The model writes the answer and calls further tools if it needs to.
What the providers document today:
| Offering | Data source | Permissions | What the assistant costs |
|---|---|---|---|
| AWS: Amazon Q Developer in the console and in chat applications | 38 APIs across seven Billing and Cost Management services, Cost Explorer among them | IAM permissions for Q and the services; read-only | included in Q Developer; free up to 25 questions with account or resource context per account a month, beyond that Pro at USD 19 per user a month |
| AWS: Billing and Cost Management MCP server, open source (Apache 2.0) since 2025-08-22 | Cost Explorer, Budgets, Cost Optimization Hub, Compute Optimizer, price lists, Storage Lens and more | the caller's credentials; the permission list includes write actions | each Cost Explorer request USD 0.01 |
| Microsoft: Azure Copilot in the portal | Cost Management, in the scope last used | everything the user can access; actions only after confirmation | no extra cost, except the Observability Agent |
| Google: Gemini Cloud Assist in Billing Reports and the FinOps hub (Preview) | configures and summarises reports in the billing console | not stated separately | not stated |
| Google: BigQuery MCP server, managed and remote | SQL on the billing export in BigQuery | MCP Tool User, BigQuery Job User, BigQuery Data Viewer roles | queries run as BigQuery jobs |
Amazon Q shows, with each answer, the APIs it called with their parameters and, where possible, a link to the matching console view (AWS Cost Explorer and AWS Budgets). Azure Copilot takes the scope last open in Cost Management and asks when it is unclear (Azure Cost Management). Gemini Cloud Assist sets a report's settings and filters in Billing Reports from a plain-language request (Google Cloud billing reports); Google says its general Cloud Assist panel is not designed for cost questions. The BigQuery MCP server limits a query to three minutes by default and a result to 3,000 rows.
Microsoft and Google also offer MCP access to their billing APIs, but it returns prices and accounts, not spend. For cost, the Azure MCP Server lists only a tool for retail prices. The MCP endpoint of the Cloud Billing API reads prices, SKUs, billing accounts and their IAM policies, and can create subaccounts, move accounts and set IAM policies.
What it means for your cost
An assistant costs money in three places. Licences: AWS lists Amazon Q Developer Pro at USD 19 per user a month, while Azure Copilot is included at no extra cost. API calls: AWS's MCP server calls the Cost Explorer API with the caller's credentials, and each request costs USD 0.01 (both list prices in US dollars with no Region stated, read on 2026-10-16); an open question can trigger several calls in a row, which is how Amazon Q describes its own approach. Tokens: every result enters the context as input and is read again at the next step (tokens, requests and context). A resource-level query over a month fills the context faster than a total by service. How agent loops multiply their cost is covered in agents and tool calls.
The gain is the speed of the first question, not a new data source: every assistant reads the same data as the console, with the same freshness and the same gaps. Google states plainly that Gemini can produce output that sounds plausible but is wrong, and recommends checking it.
Where it goes wrong
- Figures without a query. An answer quotes an amount that no tool call and no report stands behind. The cause: the model wrote the figure instead of retrieving it. The displayed API calls or the generated report show whether it was retrieved; a figure with neither does not count.
- More access than the question needs. An MCP server runs with a profile that can read the whole organisation's cost and even change resources. The cause: the server uses the caller's credentials, and the documented permission list of AWS's server includes write actions. A dedicated role limited to cost data bounds what a tool call can reach (access and roles for cost data).
- A member account taken for the organisation. The assistant reports "the company's" cost but sees one account or scope. The cause: Amazon Q aggregates member accounts automatically only when signed in to the organisation's management account (otherwise through a custom billing view), and Azure Copilot takes the scope last used.
- Cost measure and period left unstated. Two answers to the same question differ. The cause: one call ran unblended, the other amortized, or the period differed. A good question names both, and a good answer shows both.
- An assistant used as a schedule. An agent asks for cost every hour through the MCP server. The cause: every question triggers billed API calls and tokens. Recurring questions are what a saved report or a budget is built for, and neither needs a model.
Sources
Read on 2026-10-16. Model Context Protocol specification and base protocol; Linux Foundation formation of the Agentic AI Foundation; AWS cost management capabilities in Amazon Q Developer, how they work, Amazon Q Developer pricing, announcement of the Billing and Cost Management MCP server, the server's README and Cost Explorer pricing; Microsoft analyze costs using Azure Copilot, Azure Copilot overview and pricing tools in the Azure MCP Server; Google Cloud Gemini Cloud Assist in Cloud Billing, BigQuery MCP server and MCP reference for the Cloud Billing API.