> For the complete documentation index, see [llms.txt](https://developer.abacum.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://developer.abacum.ai/guides/connect-ai-to-abacum.md).

# Connect AI to Abacum

Let an AI assistant such as Claude or ChatGPT read your Abacum data with the Abacum MCP server.

The Abacum MCP server lets an AI assistant, such as Claude or ChatGPT, read the Abacum data that you can see. You ask a question in your assistant, and it gets the answer from Abacum. You do not need to open Abacum.

This is not the docs MCP server. That one reads this documentation. See [Use the docs with AI tools](/resources/use-the-docs-with-ai-tools.md).

## Before you start

* MCP is a paid feature. Ask your customer success manager to turn it on for your organization.
* A Super Admin turns the server on in **Settings** > **Connections & API**.
* You need an AI assistant plan that allows custom connectors.
* Admins, Super Admins and Members can connect. What you can read depends on your role.

## Connect your assistant

The server address is `https://mcp.abacum.ai`.

Add it as a custom connector in your assistant. Sign in with your Abacum account when the assistant asks. Single sign-on works. You do not enter a client ID or a client secret, because the assistant registers itself.

**Claude.** Open the connector settings, add a custom connector, paste the server address and sign in.

**ChatGPT.** Turn on Developer mode, create an app, paste the server address, choose OAuth and leave the client ID and secret empty. Then sign in.

**Other assistants.** Abacum supports Claude and ChatGPT. Other assistants that support remote MCP servers with OAuth 2.0 and dynamic client registration can also connect.

**Gemini Enterprise.** Gemini Enterprise does not register itself. You paste six values into Google Cloud: the server address, the authorization address, the token address, the scope, a client ID and a client secret. Turn on PKCE. Treat the client secret like a password. It is shown once.

## What the assistant can see

The assistant acts as you. It sees only what you see in Abacum. Each connection uses the permissions of the signed-in user. To limit what the assistant can read, connect with a user whose role fits the task.

| Role                  | What the assistant can read                                                                     |
| --------------------- | ----------------------------------------------------------------------------------------------- |
| Admin and Super Admin | Models, variables, scenarios, datasets, model logic, variance analysis, Space tables and charts |
| Member                | Space tables and charts in the Spaces that the member can open                                  |

## What it can do

The server reads data. Your assistant picks the tool, so you can ask in plain words.

| Tool                                                                                       | What it does                                |
| ------------------------------------------------------------------------------------------ | ------------------------------------------- |
| `abacum_get_models`, `abacum_get_scenarios`, `abacum_get_variables`, `abacum_get_datasets` | List what you have                          |
| `abacum_get_model_logic`                                                                   | Read the formulas and structure of a model  |
| `abacum_get_variance_analysis`                                                             | Compare actuals with budget or forecast     |
| `abacum_list_spaces`, `abacum_list_space_tables`, `abacum_get_space_table`                 | Find and read Space tables                  |
| `abacum_search`                                                                            | Search by name                              |
| `abacum_describe_data`, `abacum_describe_dataset`                                          | Load data so the assistant can analyze it   |
| `abacum_execute_sql`                                                                       | Run a read-only `SELECT` on the loaded data |
| `abacum_list_tables`, `abacum_unload_tables`                                               | See and free the loaded data                |

## Limits

* A SQL query returns at most 10,000 rows, and a query can be at most 10,000 characters.
* The assistant can keep 40 loaded tables at a time. Loaded data is dropped after 30 minutes without use.
* A variance analysis covers at most 12 months and 10 variables.
* A request to Abacum times out after 120 seconds.
* The normal API rate limits apply. If the assistant gets a rate limit answer, wait and ask again.
* Large models can reach the context limit of your assistant. Ask for one model or one period at a time.

## Your data and privacy

The Abacum MCP server does not run an AI model, and it does not store your prompts or the answers. It records the name of the tool and the assistant's user agent for each call.

The data that a tool returns goes to the AI provider that you connected. Treat it like data that you paste into that tool, and check your provider's terms. Tools return data as it appears in Abacum. Review the data terms of your AI provider before you connect.

## Turn it off

* A Super Admin turns the server off in the same Settings section. Access ends for everyone in the organization at once, and each person must connect again when it is turned back on. The setting applies to the whole organization.
* When Abacum suspends a user, that user's assistant loses access.
* If you work in more than one organization, connect again after you switch organization. A connection works in one organization only.