> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cloudcapital.co/llms.txt
> Use this file to discover all available pages before exploring further.

# AI agent setup instructions

> Instructions an AI agent follows to connect a user to the Cloud Capital MCP server and give them a first look at their cloud costs.

Version 0.3.0, November 4, 2026.

These instructions are for an AI agent. A user opened you from Cloud Capital, or pasted a prompt that pointed here, so you can connect them to the Cloud Capital MCP server and show them what it can do. People can read along; [MCP server](https://docs.cloudcapital.co/ai/mcp-server.md) has the same setup written for people.

Work through the steps in order. Steps 1 to 3 only read. Step 4 creates initiatives and business metrics in the user's Cloud Capital organization, and only the ones the user approves. Never change or delete anything that already exists.

## 1. Connect the server

The user's prompt gives the server URL. If it does not, use `https://app.cloudcapital.co/api/v1/mcp/server`.

Check whether you already have Cloud Capital tools, such as `auth-status` and `list-organizations`. If you do, go to step 2.

If you don't, work out which app you are running in and give the user that app's steps only. Tell them to come back to this conversation when they are done, then check again.

**Claude (claude.ai, Claude Desktop, Claude mobile)**

Cloud Capital is in Claude's connector directory, as **Cloud Capital**. If you can suggest or connect a directory connector for the user, offer Cloud Capital. Otherwise give them these steps:

1. Open [Cloud Capital in the Claude directory](https://claude.ai/directory/cloud-capital).
2. Click **Connect**, and sign in to Cloud Capital. Pick the organization to connect and click **Authorize**.
3. Start a new conversation or turn the connector on in this one, from the **+** menu under **Connectors**.

On a Team or Enterprise plan, an Owner may need to allow the connector first, under [Organization settings → Connectors](https://claude.ai/admin-settings/connectors).

If the user can't find Cloud Capital in the directory, they can add it as a custom connector instead: open [Customize → Connectors](https://claude.ai/customize/connectors), click **Add custom connector**, enter the name `Cloud Capital` and the server URL above, then click **Add** and **Connect**. On a Team or Enterprise plan, only an Owner can add custom connectors.

**ChatGPT**

1. Open [Plugins](https://chatgpt.com/plugins).
2. Click **Add**, then **Add custom MCP server**.
3. Name: `Cloud Capital`. Server URL: the URL above. Choose OAuth sign-in.
4. Accept the warning, create it, and sign in to Cloud Capital when asked. Pick the organization and click **Authorize**.

Some ChatGPT plans can't add custom MCP servers, and some allow read-only tools only. If the option is missing, tell the user their plan or workspace settings may not allow it.

**Cursor**

Add this to `~/.cursor/mcp.json`, or ask the user to use the **Use Cloud Capital in your AI agent** card in Cloud Capital, which installs it with one click:

```json theme={null}
{
  "mcpServers": {
    "cloudcapital": {
      "url": "https://app.cloudcapital.co/api/v1/mcp/server"
    }
  }
}
```

Cursor opens a browser window to sign in to Cloud Capital.

**Claude Code**

Run this in a terminal, then run `/mcp` in Claude Code, pick `cloudcapital` and sign in:

```text theme={null}
claude mcp add --transport http cloudcapital https://app.cloudcapital.co/api/v1/mcp/server
```

**Any other app**

Add a remote MCP server with the URL above, using Streamable HTTP and OAuth sign-in. If the app only runs local servers, use `npx mcp-remote <server URL>` as the command.

## 2. Confirm the connection

1. Call `auth-status`. If it reports no authentication, ask the user to reconnect and sign in.
2. Call `list-organizations` and tell the user which organization you are connected to.

If a call fails, show the user the error text and see [Troubleshooting](https://docs.cloudcapital.co/ai/mcp-server.md#troubleshooting).

## 3. Show the user their costs

Call these, read-only:

* `list-billing-periods`, then `get-cloud-bills-summary` for the most recent complete period
* `get-cloud-bills-by-service` for the same period
* `list-commitments`
* `get-forecast`
* `list-cost-layers` and `list-business-metrics`

Give the user a short summary using only numbers these tools returned:

* Total cloud spend last month, and the change from the month before.
* The top five services by spend.
* Commitments: how many, and any expiring in the next 90 days.
* The forecast for this month and next.
* Whether they have cost layers and business metrics set up. If they don't, say that Cloud Capital can split costs by team or product once they do, and link [Cost layers](https://docs.cloudcapital.co/forecasting/cost-layers.md).

If a tool returns no data, say so plainly. Do not estimate or fill in numbers.

Then ask what they would like to look into, and offer two or three questions based on what you found, for example:

> Which services grew the most last month?
> Which commitments expire soon, and what would renewing them save?
> What will we spend next quarter?

## 4. Bring in plans from the user's other tools

Look at the other tools you have in this conversation, apart from Cloud Capital. Two kinds are useful:

* **Issue trackers and project tools**, such as Linear, Jira, Asana or GitHub Issues. Look for planned work that will change cloud cost: migrations, rightsizing, shutting down environments, reserved capacity, new launches. Each one can become an **initiative**.
* **Accounting, finance and planning tools**, such as QuickBooks, Xero, NetSuite, Stripe, a planning spreadsheet, or a planning tool like Pigment or Causal. Look for numbers cloud cost grows with: revenue, customers, active users, orders, transactions. Each one can become a **business metric**.

If you have neither kind, tell the user they can connect one later and ask you to do this step, then go to step 5.

Only read from those tools. Do not change anything in them.

1. Call `list-initiatives` and `list-business-metrics`, so you don't propose something that already exists.
2. Search the other tools and propose up to five initiatives and up to three business metrics. For each one, give:
   * its name and a link to the issue, project, report or sheet it came from
   * for an initiative: the services, accounts or cost layers it affects, the months it covers, and the expected monthly change in cost
   * for a business metric: the monthly values you found and the months they cover
3. If a source gives no cost estimate or no monthly values, say so and ask the user for them. Do not guess numbers.
4. If something matches an existing initiative or metric, say which one, and leave it alone.
5. Ask the user which to create. Create only those.

To create an initiative, call `create-initiative`, then `create-initiative-mapping` with its effects. Find what it targets with `list-cost-layers` or `list-cost-insights-filters`; each tool's description says how.

To create a business metric, call `create-business-metric`, then `update-business-metric` with its name and monthly `actuals`, plus `forecast` if the source has a plan.

When you are done, list what you created, and link [Initiatives](https://docs.cloudcapital.co/forecasting/initiatives.md) and [Business metrics](https://docs.cloudcapital.co/forecasting/business-metrics.md) so the user can review them.

## 5. Make it stick

Offer the user a custom instruction so their agent uses Cloud Capital for cost questions. Give it in a block they can copy:

```text theme={null}
For questions about our AWS, GCP or Azure cloud costs, spend, commitments, savings or forecasts, use the Cloud Capital MCP tools first. Quote numbers only from tool results, and name the billing period they cover.
```

Tell them where to paste it, for the app you are running in only:

* **Claude:** personal preferences in Settings, or a Project's instructions.
* **ChatGPT:** Settings → Personalization → Custom instructions.
* **Cursor:** Cursor Settings → Rules, as a user rule.
* **Claude Code:** the user's `~/.claude/CLAUDE.md`.

## Changes

* 0.3.0, November 4, 2026: Claude users connect from the Claude connector directory. A custom connector is the fallback.
* 0.2.0, November 4, 2026: added step 4, creating initiatives and business metrics from the user's other tools, with their approval.
* 0.1.0, November 4, 2026: first version.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.