# AI access over MCP

> AmunCore is an MCP (Model Context Protocol) server. Claude, ChatGPT and any MCP client can list your applications and endpoints, read and write through them, work out totals, find records and read your documents, with the same permissions, row rules and audit log as the REST API.

## Connect an assistant

The MCP address is `https://amuncore.com/mcp` (on your own servers, your own address followed by `/mcp`). An admin first switches on AI access on the AI access page.

1. OAuth connector (recommended, no setup): in Claude (desktop, web or mobile) open Settings, Connectors, Add custom connector, enter the MCP address and press Connect. Sign in to AmunCore and press Allow. The assistant then works as that person, with their own access.
2. Connection token: for ChatGPT, scripts or any HTTP client, create a connection on the AI access page and send its token in the `MCP-Token` header with each request.
3. Older Claude Desktop versions can use the mcp-remote bridge (needs Node.js); the AI access page shows the exact configuration.

MCP clients discover sign-in through [the OAuth protected-resource metadata](https://amuncore.com/.well-known/oauth-protected-resource).

## Tools

Assistants discover tools with `tools/list` and call them with `tools/call`. Each tool declares its parameters and returns structured JSON.

| Tool | What it does |
|---|---|
| list_applications | List the company's applications |
| get_application | Details of one application |
| list_endpoints | List the endpoints of an application, with their descriptions |
| call_api | Read or change data through any allowed endpoint (GET, POST, PUT, DELETE) |
| get_totals | How many, total, average, lowest, highest or how many different values, by a field or by month, day or year, with conditions, worked out by the database over every record (read only) |
| find_records | Records with conditions and in an order, for example the 10 largest unpaid invoices, chosen among every record, with how many match (read only) |
| list_files | List stored documents |
| get_file_text | Read a document's text |
| search_files | Search documents by name and content |
| get_dashboard_stats | Overview figures |
| create_application | Register a new application or database |
| create_endpoint | Add an endpoint to an application |
| toggle_endpoint | Switch an endpoint on or off |
| regenerate_api_key | Make a new API key (the old one stops working) |
| get_audit_logs | Read recent API and AI activity |
| get_license_status | Check the plan and its limits |
| check_plan_limits | Check whether more can be created |

The company chooses which tools its assistants may use. For everyday use the read tools plus call_api are recommended; get_totals and find_records are off until ticked.

## What each connection can be limited to

- Applications and endpoints: all, several or one.
- Tools: exactly which tools are offered.
- Read-only: write requests are refused.
- Who can use it: everyone in the company or chosen people.
- Expiry date, calls per minute and allowed IP addresses.
- Active or disabled, without deleting it.

## Safety

- Every tool call goes through the endpoint as the caller: permissions, row rules, column rules, approval policies and fixed filters apply exactly as for a REST call.
- Writes can require named fields first and a human approval before anything is saved.
- Text from records and documents is treated as data, never as instructions.
- Every call is recorded in the audit log, and AI access can be switched off for the whole company at once.

## Related

- [Client Connectors](https://amuncore.com/docs/client-connectors.md): the same for your own clients' assistants.
- [AI agents](https://amuncore.com/docs/agents.md): AmunCore's own AI agents.
- [AmunCore overview](https://amuncore.com/llms.txt)
