Cloudeval AI MCP server

Cloudeval CLI

Your cloud, in the terminal: evaluated, reported, and agent-ready.

Cloudeval CLI turns ARM templates, GitHub-hosted IaC, and live Azure context into cost, architecture, and Well-Architected signals. Use it as a terminal UI, a scriptable automation client, or an MCP server for Codex, Cursor, Claude, VS Code, and any stdio JSON-RPC client.

Why Use It

For	What you get
Terminal users	A full TUI with chat, Agent mode, workspace tabs, thread switching, slash commands, a context rail, task ledger, artifact chips, and local SQLite session history.
Automation	Stable json, ndjson, markdown, and text output; payloads on stdout; progress and prompts on stderr; predictable exit codes.
Agents and CI	Scoped access-key credentials, redacted output by default, MCP toolsets, recipes, and machine-readable capability metadata.

Install

Node.js 20+ users can install from npm:

npm install -g @ganakailabs/cloudeval-cli
cloudeval --help

macOS, Linux, WSL2, Git Bash, and PowerShell 7+ on Windows or Linux can use the standalone release installer:

curl -fsSL https://cli.cloudeval.ai/install.sh | bash

irm https://cli.cloudeval.ai/install.ps1 | iex

Then reload your shell and sign in:

source ~/.bashrc   # or: source ~/.zshrc
cloudeval login
cloudeval status
cloudeval chat

Device login goes through cloudeval.ai and always asks the browser auth provider to show the account chooser, so users can choose the intended work email even when another account is already signed in. No local Azure app registration is needed for normal CLI use.

The installer:

• downloads checksum-verified GitHub release assets and installs cloudeval;
• creates the eva and cloud aliases on non-Windows platforms;
• can install shell completions for bash, zsh, and fish;
• can offer concise MCP setup for detected Codex, Claude Desktop, Cursor, and VS Code clients, skipping clients where Cloudeval MCP is already configured and avoiding prompts when only manual-only setup remains;
• asks whether to share limited CLI telemetry, defaulting to yes; declining writes telemetry.enabled=false;
• explains credential setup but does not create access keys or write secrets into MCP client config;
• shows compact labeled progress bars in interactive terminals;
• uses connect/stall timeouts so slow CDN transfers fail clearly.

Useful controls:

curl -fsSL https://cli.cloudeval.ai/install.sh | CLOUDEVAL_INSTALL_AGENT_SETUP=0 bash
curl -fsSL https://cli.cloudeval.ai/install.sh | CLOUDEVAL_INSTALL_MCP_CLIENTS=codex,cursor bash
curl -fsSL https://cli.cloudeval.ai/install.sh | CLOUDEVAL_TELEMETRY=0 bash

$env:CLOUDEVAL_ASSUME_YES = "1"
irm https://cli.cloudeval.ai/install.ps1 | iex

The bash installer can also detect agent clients and offer MCP setup. The PowerShell installer installs the verified binary, yoga.wasm, license notices, PATH, and optional PowerShell tab completions. Run cloudeval mcp setup afterward when you want MCP client configuration.

Telemetry

Cloudeval CLI sends curated custom events to Azure Application Insights by default. Events cover command family, success, duration, safe option enums, CLI version, Node/runtime version, OS major version, architecture, install source, update/install outcomes, MCP tool names, and TUI launch/exit metadata. After login, events may include the signed-in email and first/last/full name.

Telemetry never sends raw prompts, command output, tokens, local paths, project or resource identifiers, account/session/tenant identifiers, cloud resource names, stack traces, or raw error messages. Disable or re-enable it anytime:

cloudeval config set telemetry.enabled false
cloudeval config get telemetry.enabled --format json
cloudeval config set telemetry.enabled true
cloudeval config unset telemetry.enabled

Environment overrides take precedence for a single run:

CLOUDEVAL_TELEMETRY=0 cloudeval status --format json
CLOUDEVAL_TELEMETRY=1 cloudeval --help

Update later with:

cloudeval update --check
cloudeval update --yes

After an update, restart or reload configured MCP clients when you are ready to load newly exposed Cloudeval tools, resources, or prompts. Cloudeval does not restart Codex, Claude, Cursor, VS Code, or other MCP hosts automatically.

Uninstall local installer-owned artifacts while keeping Cloudeval config, sessions, and auth by default:

cloudeval uninstall --dry-run
cloudeval uninstall --yes
cloudeval uninstall --yes --remove-config  # also removes ~/.config/cloudeval
npm uninstall -g @ganakailabs/cloudeval-cli # if installed through npm

Start Here

cloudeval                         # Terminal UI
cloudeval tui --graph-diagram ascii
cloudeval ask "Summarize my cloud risk" --format json
cloudeval agent "Find cost and architecture risks" --format json
cloudeval agents list
cloudeval agents run cost --project <project-id> --format json
cloudeval recipes list
cloudeval projects list
cloudeval uninstall --dry-run
cloudeval projects graph insights <project-id> --focus impact --resource <resource-id> --format json
cloudeval validate template --template-file template.json --parameters-file parameters.json --rule <check-id> --details --wait --progress stderr --wait-timeout 600000 --format json
cloudeval validate tests --template-file template.json --parameters-file parameters.json --wait --progress stderr --wait-timeout 600000 --format json
cloudeval rules search "public network" --format json
cloudeval reports list
cloudeval actions list --type architecture,cost,unit-tests --format json
cloudeval actions open --print-url --no-open
cloudeval review --repo owner/repo --ref feature/infra-change --commit-sha <sha> --github-checks --sarif --output cloudeval-review --format json --non-interactive
cloudeval capabilities --format json
cloudeval doctor --deep

Full docs: Get started with the CLI and CLI command reference.

Inside the Terminal UI, use the Thread control or /thread to switch open chat sessions, recent Cloudeval chat threads, and local CLI sessions. /thread new starts another independent open session, and /open jumps to the same chat thread in Cloudeval when the active session has a thread id. Roomy terminals show a context rail with project, thread, model, mode, profile, report artifact chips; narrower terminals keep the chat first and expose the same controls through the composer and slash commands. Typing / opens a bottom command completion strip; use Tab or Up/Down to move, Right to accept the ghost text, and Enter to choose the highlighted command. Streaming work appears as a task ledger in the thread, and the bottom composer stays docked so prompt entry does not compete with the transcript. Grounded answers show numbered citations and a Sources section instead of raw [S_tool_...] tags, with citation numbers highlighted inline; /copy copies the latest assistant response and /download writes a Markdown transcript with the same references. Graph insight blocks render as bordered terminal cards instead of exposing raw graph-insight markers; when a card contains a conservative Mermaid flowchart, --graph-diagram auto renders a terminal diagram in roomy TTYs, unicode or ascii force a mode, and off keeps the Mermaid source fallback. Unsupported Mermaid syntax stays visible as source instead of breaking the transcript. Negotiated chat visualizations render directly in the TUI: line/area trends use Unicode plots; bar, column, histogram, pie, doughnut, radar, and polar data use width-aware bars; scatter and heatmap data use compact terminal grids; unsupported chart families use the artifact's table fallback. Mermaid flow edges render as an edge list, with bounded Mermaid source as the fallback when no edges can be extracted. HITL approval prompts require an explicit option selection or typed answer; pressing Enter on a blank approval prompt does not choose the recommended option. Project and Connection tabs show a selected-item detail pane for backend fields, report coverage, sync state, and linked records; use J/K or Up/Down on Projects and Connections to move the selected row, then Enter to confirm it. The billing header separates credits left from observed credits used so usage does not look like the current budget. Use the Profile control or /profile cost to run the current prompt with an Agent Profile; selecting a profile switches the TUI to Agent mode, and selecting Ask mode clears the profile back to the default chat flow. Starter prompts stay hidden until you run /starter. Press Esc from the prompt to leave text editing so tab, arrow, and number shortcuts move through controls and tabs; type again to resume editing. Busy loaders and the input cursor can be disabled with --no-anim. The banner details include the logged-in user. Focused controls and the active top tab use the shared warm banner-yellow accent, with the active tab filled across its full button interior.

The CLI advertises cloudeval.visualization/v1, flint-v1, and mermaid-v11 capabilities on chat requests. The backend compiles chart intent; the CLI validates the bounded artifact and renders terminal-safe output without a browser or native SVG helper. ask and agent JSON results include data.visualizations when present, and NDJSON emits a visualization event as well as including the artifacts in the final result. Final JSON/NDJSON responses, Markdown output, and local conversation history retain validated artifact fences even when streamed prose omits or corrupts the chart payload. Text output remains the live prose stream. See the visualization artifact contract.

Core Workflows

Goal	Terminal UI	Script or CI	MCP
Grounded cloud chat	cloudeval or cloudeval chat	cloudeval ask "..." --format json	ask
Deeper analysis	Agent mode in the TUI	cloudeval agent "..." --format json	planner-style tool flows
Agent Profiles	TUI Profile control and Chat picker	cloudeval agents list/show/run	agent_profiles_* tools
Reusable workflow	prompt suggestions	cloudeval recipes list/show/run	recipes_* tools
Projects and reports	workspace panels	projects, reports, open	projects_*, reports_*
Issues	/app/issues	issues list/get/open	n/a (use CLI; MCP has report/deeplink tools)
Graph intelligence	project graph views	projects graph ...	projects_graph_* tools
Template validation	n/a	validate, rules	template_*, rules_*
Billing	billing panel and links	billing, credits	billing_* toolset
Automation discovery	n/a	capabilities --format json	capabilities_get

Agent Profile ids include architecture, cost, triage, remediation, visual-explainer, scripter, change-reviewer, evidence-auditor, and security-reviewer. Display names may contain spaces. The Architecture profile includes the Well-Architected review lens, so there is no separate Well-Architected Agent Profile. When agents run omits a prompt, the CLI uses a starter prompt for the selected project source and profile mode: template or live sync, ask or agent. The choice is deterministic for automation. Profile runs send only agent_profile_id; Cloudeval applies profile instructions, planning lens, and response defaults on the backend. agents list and agents show first try the backend profile catalog; if the profile catalog endpoint requires sign-in or is not available, they fall back to the bundled public catalog so discovery still works. agents run still requires authenticated backend access. In the TUI, the Profile selector uses the same canonical IDs and sends the selected agent_profile_id with chat streams.

Run cloudeval <command> --help for exact flags.

Access Keys For CI And Agents

Use cloudeval login for humans. The browser approval page requests an account chooser on every login. Use scoped access keys for CI, hosted agents, and long-running automation.

Stored device-login sessions refresh automatically before authenticated requests. If the TUI or cloudeval ask receives an expired-token response from the chat stream, the CLI refreshes the stored session and retries that request once. If the refresh token is revoked or expired, run cloudeval login again.

Create an access key after login and project selection:

cloudeval projects list
cloudeval credentials templates --format json
cloudeval credentials create \
  --template ci \
  --name github-actions-prod \
  --project <project-id> \
  --expires 90d \
  --idempotency-key "$(uuidgen)" \
  --format github-actions

--format github-actions prints CLOUDEVAL_ACCESS_KEY and CLOUDEVAL_PROJECT_ID once. The raw key is not shown again by credentials list or credentials inspect.

Test a scoped access key without putting it in shell history:

printf '%s\n' "$CLOUDEVAL_ACCESS_KEY" | cloudeval projects list \
  --access-key-stdin \
  --format json \
  --non-interactive

Credential rules:

• prefer --access-key-stdin or CLOUDEVAL_ACCESS_KEY;
• --access-key is accepted but warns because process arguments and shell history can leak;
• old beta names --api-key, --api-key-stdin, and CLOUDEVAL_API_KEY fail with a migration error;
• access-key-shaped strings, authorization headers, and sensitive URL query parameters are redacted by default;
• credential create output files are written with private permissions on POSIX systems.

MCP For Coding Agents

Start MCP after signing in, or provide a scoped CLOUDEVAL_ACCESS_KEY in the host environment:

cloudeval login
cloudeval mcp serve
cloudeval mcp serve --toolset readonly

Client setup examples:

codex mcp add cloudeval -- cloudeval mcp serve --toolset readonly
cloudeval mcp setup cursor --dry-run --toolset reports --format json
cloudeval mcp setup vscode --dry-run --toolset readonly --format json

MCP rules:

• tool names use underscores such as projects_list, recipes_list, and billing_summary;
• dotted tool names remain compatibility aliases;
• stdout is JSON-RPC only and [cloudeval-mcp] diagnostics go to stderr;
• MCP tool schemas do not accept per-call access-key arguments;
• mcp serve does not support --access-key-stdin because stdin is the protocol stream.
• readonly includes safe inspection tools for projects, reports, billing, connections, credentials, config, models, sessions, auth, status, doctor, and recipes; generation, downloads, checkouts, credential mutation, browser opens, and diagram file writes stay explicit.

For billing inspection, use billing_ledger for individual usage attempts and credit charges, billing_usage for aggregates, and billing_summary for current entitlement. Ledger filters default to 30 calendar days; startAt and endAt override their corresponding range bounds. Pass data.next_cursor back as cursor with the same filters while data.has_more is true. Ledger page size defaults to 25 and is clamped to 1–100.

billing_invoices returns subscription invoices, paid top-up history and billing-cycle status. Fetching this data can create missing provider invoice records for already-paid top-ups and persist receipt links. It therefore requires explicit --toolset billing or --toolset all selection and is excluded from readonly. Its result limit defaults to 25, is clamped to 1–50 per collection, and has no pagination cursor. These tools require billing read access through the server's configured credential; they do not initiate a purchase or change the subscription.

Developer setup details: cli.cloudeval.ai/developer/.

Recipes And Skills

Cloudeval recipes are reusable workflows for agents and humans. Current recipes cover cost review, WAF triage, architecture review, template project review, report summaries, report generation planning, report export packs, billing review, top-up readiness, project inventory and healthchecks, connection audit, credential setup and rotation, model selection, session recovery, CLI onboarding checks, frontend workspace links, architecture/dependency diagram exports, and MCP setup.

cloudeval recipes list
cloudeval recipes show cloudeval-cloud-cost-review
cloudeval recipes run cloudeval-cloud-cost-review --project <project-id> --format json --non-interactive
cloudeval recipes show cloudeval-architecture-diagram-export
cloudeval recipes run cloudeval-dependency-diagram-export --project <project-id> --output-path ./dependency.svg

Ask/agent-backed recipes may consume model credits. Recipes that would create projects, write report or diagram files, change MCP config, mutate credentials, open browsers, or start checkout flows print explicit commands instead of performing those side effects implicitly. Portable agent instructions live under skills/; MCP remains the preferred execution path for Codex, Cursor, Claude, and other agents.

Project Example

curl -L -o template.json \
  https://raw.githubusercontent.com/Azure/azure-quickstart-templates/master/quickstarts/microsoft.compute/1vm-2nics-2subnets-1vnet/azuredeploy.json

cloudeval projects create \
  --name "Azure VM network review" \
  --provider azure \
  --template-file ./template.json \
  --format json

Use --template-url when you do not want a local file. Follow with reports run, reports download, and projects export-diagram as needed.

Output, Auth, And Privacy

cloudeval login
cloudeval login --headless
cloudeval auth status
cloudeval auth status --show-sensitive-ids
cloudeval help agents
cloudeval agents list

Output contract:

• cloudeval login opens or prints a cloudeval.ai/device/login approval URL with an account chooser hint for the web auth provider;
• machine-readable commands write payloads to stdout;
• prompts, progress, browser-open messages, and warnings go to stderr;
• ask and agent support --progress none, --quiet, or --format ndjson --progress ndjson;
• validate template and validate tests support --progress stderr or --progress ndjson with --wait; validation progress always goes to stderr so final JSON/NDJSON remains parseable on stdout. Completed progress includes 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;
• with --non-interactive, human approval exits with code 6 and returns HITL_REQUIRED;
• interactive HITL prompts require an explicit option number, yes/no-style answer, or typed response; blank Enter does not approve the recommended option;
• --show-sensitive-ids shows full account/session-style IDs only on trusted machines. It does not unredact tokens.

Docs

Link	Purpose
Get started with the CLI	Install, login, create a project, and ask a grounded question
CLI command reference	Full command and flag list
Terminal UI	TUI navigation and keyboard model
MCP client setup	Codex, Cursor, Claude, VS Code, and generic MCP hosts
Agent behavior and automation safety	Safe automation conventions
Troubleshooting	Sign-in, onboarding, reports, and billing