Where it plugs in

# Where it plugs in

The Understanding API is one service with three doors: a REST endpoint, a remote MCP server, and function definitions for your own agent. This page lists where each door fits, the address to use, and how you sign in. Each service has its own page with the exact steps.

Not open yet. Opens with early access.

## Pick your service

| You are using | Address | Sign in with | Page |
| --- | --- | --- | --- |
| A banking product, a card program or a back-office tool | `POST https://api.theunderstanding.company/v1/understand`, and `/v1/understand/batch` | An API key | [Banking apps](/api/docs/integrations/banking-apps) |
| ChatGPT | `https://api.theunderstanding.company/mcp/chatgpt` | Sign in (OAuth) | [ChatGPT](/api/docs/integrations/chatgpt) |
| Claude Desktop, claude.ai or Claude Code | `https://api.theunderstanding.company/mcp/claude` | Sign in. Claude Code can also use a key. | [Claude](/api/docs/integrations/claude) |
| Grok Bot | `https://api.theunderstanding.company/mcp/cursor` | Sign in when the app asks, or a key if it offers a header | [Grok Bot](/api/docs/integrations/grok-bot) |
| Grok Build and the Grok chat app | `https://api.theunderstanding.company/mcp/grok` | A key in a header (Grok Build). The sign-in Grok shows (Grok chat). | [Grok Bot](/api/docs/integrations/grok-bot) |
| Meta's Muse, Muse Code, Meta Model API | `https://api.theunderstanding.company/mcp/muse`, or function definitions | A key | [Muse](/api/docs/integrations/muse) |
| Instinct | `https://api.theunderstanding.company/mcp` | A key (settings to try) | [Instinct](/api/docs/integrations/instinct) |
| Cursor and VS Code | `https://api.theunderstanding.company/mcp/cursor` for Cursor, `https://api.theunderstanding.company/mcp` for VS Code | A key in a header, or sign in | [Cursor and VS Code](/api/docs/integrations/cursor-vscode) |
| Your own agent on OpenAI's API | `https://api.theunderstanding.company/mcp/chatgpt`, or `GET /v1/tools?format=openai` | A key | [OpenAI API](/api/docs/integrations/openai-api) |
| Your own agent on Anthropic's API | `https://api.theunderstanding.company/mcp/claude`, or `GET /v1/tools?format=anthropic` | A key | [Anthropic API](/api/docs/integrations/anthropic-api) |
| Your own agent on xAI's API | `https://api.theunderstanding.company/mcp/grok`, or `GET /v1/tools?format=xai` | A key | [xAI API](/api/docs/integrations/xai-api) |

Anything else that speaks MCP uses `https://api.theunderstanding.company/mcp` and a key. See [Connect via MCP](/api/docs/guides/mcp).

## What is the same everywhere

- The questions. Is this merchant a restricted one, is this spending essential or discretionary, which charge does this receipt belong to, which transaction is the refund for this purchase, and your own yes, pick-one or scored question. Also: can this purchase still be returned, which support queue fits a message, and three health questions that take derived numbers only. The eleven tools are `check_restricted_merchant`, `classify_spend_type`, `match_receipt`, `pair_refund`, `check_return_window`, `route_support_message`, `score_signal_baseline`, `route_member_message`, `check_document_complete`, `ask_closed_question` and `get_usage`. The three health tools are offered at the standard, Claude and Cursor addresses and not at the ChatGPT, Grok or Muse ones.
- The answer. A closed answer, the model's probability, reason codes and an abstain flag. When `abstain` is true, a person decides. The probability is not calibrated.
- Read-only. No tool on your key moves money or changes anything. Each is marked read-only and idempotent. (Card controls exist only for our own keys: [Cards and balances](/api/docs/use#cards).)
- The price and the limits. A call costs the same through any door, and counts toward the same limit of 60 requests a minute per key. See [Pricing](/api/docs/pricing) and [Rate limits](/api/docs/limits#rate-limits).
- The input checks. A request that holds a full card number, a Social Security number or an account number with its routing number is refused with a 400 before it reaches the model.

## What differs

Each assistant reads tool descriptions and results differently, so the MCP server has a tuned address for each: wording, the shape of the result, and which tools are offered. A tuned address never changes what a tool does, what it costs or who can call it. Instinct gets the standard server, because it publishes nothing to tune against. The ChatGPT, Grok and Muse addresses offer the banking tools only: those services have their own rules about health data. The table is on [Connect via MCP](/api/docs/guides/mcp).

## Not sure which door?

If you are building a product, start with [Banking apps](/api/docs/integrations/banking-apps) and the [Quickstart](/api/docs/quickstart). If you want an assistant you already use to answer questions about your own statements, start with its page above. If you build agents on a model API, use the page for that API.

Source: /api/docs/integrations