MCP Overview

LevelFour runs a remote Model Context Protocol server at https://mcp.levelfour.ai/mcp. Point Claude Code, Claude, Cursor, VS Code, Windsurf or any other MCP client at it, and the model can read your cloud spend, the savings opportunities LevelFour has found, and what you have already booked. It serves the same data as the dashboard and the REST API, shaped so a model can ask for it in one call instead of five.

Your organization needs the MCP module, and only an organization administrator can approve a connection. Without the module the approval screen says so and nothing completes. Talk to your LevelFour contact.

Connect

Add the server

claude mcp add --transport http levelfour https://mcp.levelfour.ai/mcp

Type the URL exactly as written, path included. A client that sends the bare host or a trailing slash will not match what the server advertises.

Run /mcp inside Claude Code and choose Authenticate. Your browser opens the LevelFour approval screen, you sign in with the account you already have, and you approve the connection for one organization.

No key is copied and nothing lands in a config file. The token expires in an hour, refreshes on its own, and works at this server and nowhere else.

Read Approving a connection before you click through it.

Confirm it worked

Ask your client to call get-identity:

{ "plan": "enterprise", "enabled_providers": ["aws"], "executions_remaining": 12, "days_remaining": 30 }

That answer is the proof: the credential is live and bound to an organization. A model answering a cost question is not proof, because it can answer badly for reasons that have nothing to do with the connection.

If no tools appear at all, restart the client. Most read their MCP config once at startup.

Ask it something

Why did our AWS bill go up last month, and what can we do about it?

Ask it this has more that work as written, and the ones that do not.

l4 mcp install is the other way in. One command writes the server into every MCP client on the machine, across Claude Code, Claude Desktop, Cursor, VS Code and Windsurf. It installs an API key rather than signing you in, so reach for it when you want five clients wired at once, or when something runs unattended.

The server

Property	Value
Endpoint	`https://mcp.levelfour.ai/mcp`
Transport	Streamable HTTP, stateless
Authorization	OAuth 2.1 with PKCE, or a LevelFour API key in `Authorization: Bearer`
Issuer	`https://mcp.levelfour.ai`
Tools	31 tools. Most read; the rest carry a savings from the decision to a rollout, in LevelFour, never in your cloud
Prompts	5
Resources	4
Rows per call	50 maximum

Every tool resolves your organization from the credential, never from an argument. If you belong to two organizations, connect twice under different names and approve each separately.

What it answers

Area	Tools
Spend	`get-cost-summary`, `get-cost-breakdown`, `get-cost-series`, `get-cost-growth`, `get-costs-by-tag`, `get-usage-costs`, `get-cost-forecast`
Savings opportunities	`list-recommendations`, `get-recommendation`, `get-savings-overview`, `get-potential-savings`
Deciding and delivering	`decide-recommendation`, `set-implementation-method`, `get-savings-readiness`, `start-rollout`
Applying it yourself	`get-manual-instructions`
Automated Savings access	`get-savings-grant`, `create-savings-grant`
Pull request delivery	`connect-version-control`, `list-repositories`
Approvals	`list-approval-requests`, `release-rollout`
Watching a rollout	`get-implementation-status`, `get-execution-logs`
Realized savings	`get-realized-savings`
Commitments	`get-commitment-overview`, `list-commitments`
Spikes	`list-anomalies`, `update-anomaly-status`
Context	`get-identity`, `list-accounts`

Potential savings and realized savings are separate quantities. Potential is unbooked opportunity, realized is applied and measured. Adding them together double counts.

The seven tools that write are spread across the rows above rather than gathered into one, because they no longer do one thing: a connection approved for changes can decide, choose how a change is delivered, clear what that format needs first, queue it, release a colleague's, and close an anomaly. None of them touches a cloud resource. See Approving a connection.

Alongside the tools the server ships 5 prompts and 4 resources. The prompts encode a full analysis in order, with the mistakes a model reliably makes already corrected. The resources tell a model which services, accounts, tags and environments your data actually contains, so it stops guessing at filter values. See Prompts and resources.

Connecting a client covers Claude Code, Claude, Cursor, VS Code and Windsurf, and the two credentials
Approving a connection is what the browser asks you, and when to refuse
Tools is the generated reference, with every parameter, enum and default
Reading a result is what comes back: the pagination envelope, truncation, limits and errors
Ask it this has questions that work as written
CLI has the write commands this server does not expose