# MCP server for AI agents

> Let Claude, ChatGPT, Cursor or your own agent work the workspace, limited to what its key allows.

- Canonical page: https://synconone.com/docs/api-mcp
- Section: Developers
- Updated: 2026-10-22

Syncon One runs an MCP (Model Context Protocol) server, so an AI agent such as Claude, ChatGPT, Cursor or an agent in your own workflow can work the workspace through the API. Every endpoint is a tool, and an agent can do only what its API key allows.

## Connect

The endpoint is `POST https://synconone.com/api/v1/mcp`, with your API key as a bearer token. It speaks the Streamable HTTP form of MCP: one JSON message in, one JSON answer out, no session. Clients that take a URL and headers use it directly. A client that only speaks stdio can go through a bridge:

```bash
npx mcp-remote https://synconone.com/api/v1/mcp --header "Authorization: Bearer sk_live_YOUR_KEY"
```

## Tools

A tool is named after its operation, such as `list_tickets`, `create_person`, `join_conversation` or `reply_to_conversation`. The list an agent sees is narrowed to the scopes its key holds, so it is not offered what it would be refused. Each tool says whether it only reads, and which deletes cannot be undone. The names and fields match the [reference](https://synconone.com/docs/api-reference).

## Keeping an agent safe

- Give the agent its own key with the fewest scopes that do the job. Read-only scopes cannot change anything.
- Text in records, notes, tickets and conversations was written by customers and other third parties. The server tells the agent to treat it as data and never as instructions; still, do not give an agent that reads untrusted text a key that can delete or send.
- A tool call is an ordinary API request with the same key, so scopes, roles, record visibility, rate limits and the record of changes all apply, and each change names the key.
- Sending a reply or scheduling a meeting reaches real people. Ask the agent to show you the wording and wait for your go-ahead.