MCP Client Setup for Cloud Architecture Review

Use Cloudeval’s MCP server when an agent should call Cloudeval as a tool server instead of parsing CLI text output. This page covers running the MCP server on your own computer. For browser chat or a hosted endpoint, see Hosted MCP with Glama, including scoped access keys, real setup screenshots, and current Claude/ChatGPT verification limits.

Cloudeval command line and agent panel showing actions for ChatGPT, Claude, Codex, Cursor, and VS Code — Cloudeval can hand the current project context to AI clients and copy setup prompts for Codex, Cursor, Claude, and VS Code.

Video walkthrough: make cloud evaluations available to agents, automation workflows, and local developer tools.

Before you configure a client

Check the local MCP surface first:

cloudeval mcp status --format json
cloudeval doctor --mcp --format json

Authenticate with one of these options:

stored cloudeval login credentials
stored cloudeval login --headless credentials for remote or SSH sessions
a scoped CLOUDEVAL_ACCESS_KEY provided through your MCP client’s private environment configuration
--machine when your environment supports machine access

mcp serve uses stdin for MCP protocol messages, so run login before starting the MCP server. A container does not inherit host login state: supply your own scoped access key at runtime. Cloudeval-hosted project and report calls require account access; the MCP server does not bundle a Cloudeval account.

Choose a client

Client-specific steps also have stable in-page anchors for deeplinks from the app (#codex, #cursor, #claude-code, #vs-code).

Codex

Generate the Codex setup guidance:

cloudeval mcp setup codex --dry-run

A typical local registration uses the Cloudeval stdio server:

codex mcp add cloudeval -- cloudeval mcp serve

Use a focused toolset when Codex should only inspect part of the product:

codex mcp add cloudeval-reports -- cloudeval mcp serve --toolset reports

Cursor

Generate Cursor setup guidance:

cloudeval mcp setup cursor --dry-run

Use the generated mcpServers entry in Cursor’s MCP configuration surface. Keep secrets in Cursor environment configuration or stored Cloudeval login state, not inline command arguments.

Claude Code

If your Claude Code environment accepts an mcpServers JSON entry, start with the generic output:

cloudeval mcp setup generic --dry-run --format json

If your local Cloudeval CLI supports a Claude-specific target for your installed Claude client, you can inspect that output too:

cloudeval mcp setup claude --dry-run

Use the generated stdio command and keep authentication in stored login state or environment variables.

VS Code

Generate VS Code MCP configuration for the current workspace:

cloudeval mcp setup vscode --dry-run

By default this targets .vscode/mcp.json in the current directory. Use the generated stdio command in your VS Code MCP extension or built-in MCP configuration surface. If your extension expects generic mcpServers JSON instead, you can still use:

cloudeval mcp setup generic --dry-run --format json

The important part is that the server command runs:

cloudeval mcp serve

mcp setup --toolset supports all, readonly, projects, reports, and billing. Use mcp serve --toolset graph or mcp serve --toolset validation when the editor agent needs those narrower surfaces.

Focus tool access

Use focused toolsets when the assistant should have a narrower Cloudeval surface.

Toolset	Use it for
`all`	Full Cloudeval MCP surface
`readonly`	Read-only project, report, billing, and capability discovery
`projects`	Project lookup, creation handoff, project deeplinks, and project context
`reports`	Report runs, report downloads, report deeplinks, and latest report context
`billing`	Billing inspection, invoice retrieval, and explicit checkout tools
`graph`	Project graph, timeline, diff, sync runs, and insight reads
`validation`	Template validation, parsing, and validation rule catalog reads

Example:

cloudeval mcp serve --toolset readonly
cloudeval mcp serve --toolset graph
cloudeval mcp serve --toolset validation

From CLI 0.38.4, billing_invoices is available only through the explicit billing or all toolsets. It retrieves subscription invoices, paid top-up history, and billing-cycle status. The service may create missing provider invoice records for already-paid top-ups and persist receipt links, so it is excluded from readonly. Its limit defaults to 25 and is clamped to 1–50 per collection; it has no pagination cursor. Use billing_ledger for individual usage attempts and credit charges, billing_usage for aggregates, and billing_summary for current entitlement. Ledger queries default to 30 calendar days. Explicit startAt and endAt ISO timestamps override the corresponding range bounds; the start is inclusive and the end is exclusive. Reuse data.next_cursor as cursor with the same filters while data.has_more is true. These tools require billing read access. Agent Profile discovery tools agent_profiles_list and agent_profiles_get are available in the full and read-only MCP surfaces. agent_profiles_run executes a profile and is excluded from readonly. They use the same canonical profile ids as the CLI: architecture, cost, triage, remediation, visual-explainer, scripter, change-reviewer, evidence-auditor, and security-reviewer. The validation toolset exposes template_validate, template_test, template_parse, rules_categories, rules_search, and rules_get. template_validate, template_test, and template_parse all accept an optional parametersFile argument. template_validate accepts ruleId for one check or ruleNames for multiple check ids, details for frontend-style per-check evidence, and wait when the client needs completed validation results instead of only a submitted job. template_test accepts includeTests, skipTests, category, testGroups, and wait. Use waitTimeoutMs with wait for bounded automation runs. When a validation call uses wait, MCP clients can include _meta.progressToken in the tools/call params to receive notifications/progress updates for queued, running, and completed states. Completed progress messages include failing check/test details such as message, recommendation, severity, and file/template or resource location when available. If a completed backend result only has a worker-local temp file path, Cloudeval reports the submitted template filename instead. The graph toolset exposes projects_graph_get, projects_graph_timeline, projects_graph_diff, projects_graph_insights, and projects_graph_sync_runs.

Verify the integration

After configuring the client:

Restart the MCP client if it requires a restart.
Ask the client to list available Cloudeval tools.
Run a low-risk command such as capability discovery or project listing.
Confirm the result uses the expected Cloudeval profile, base URL, and auth mode.

Cloudeval agent result explaining selected Azure resources with diagram context and generated Mermaid output — A useful agent result should stay grounded in Cloudeval project context, selected resources, and the visible diagram or report state.

Cloudeval AI is also published in the official MCP Registry as io.github.ganakailabs/cloudeval. The listing describes the installable npm/stdio server and read-only access-key configuration; the registry does not host your server or provide account access. You can also find Cloudeval AI in the Glama MCP directory. For the optional hosted route, follow Hosted MCP with Glama. Its listing and build checks do not provide a Cloudeval account or verify access to your projects.

Common mistakes

Putting an access key in a Dockerfile, build step, or committed client config. Keep it in private runtime environment configuration.
Calling billing_invoices from readonly on CLI 0.38.4 or later. Select billing explicitly when you intend to retrieve receipts, then restart the MCP client so it refreshes the tool list.

Next step

Use Agent and automation rules for safety defaults, or CLI command reference for the MCP command surface.

MCP tools for Azure infrastructure review — choose Azure MCP, Bicep MCP, or Cloudeval MCP by review job.

​Before you configure a client

​Choose a client

​Codex

​Cursor

​Claude Code

​VS Code

​Focus tool access

​Verify the integration

​Common mistakes

​Next step