Using SafeMeet from an AI assistant
The hosted MCP server and REST API, how an assistant signs in, and what it may do.
Connect
SafeMeet has a hosted MCP server at https://app.safemeet.now/api/mcp and a
REST API at https://app.safemeet.now/api/v1. Both cover every feature of the
app: finding a chat, reading and replying, leaving, reporting or blocking, and
your account settings.
- Claude, ChatGPT, Cursor, VS Code and other MCP clients: add a custom connector or remote MCP server with the URL above.
- Claude Code:
claude mcp add --transport http SafeMeet https://app.safemeet.now/api/mcp - Codex:
codex mcp add SafeMeet --url https://app.safemeet.now/api/mcp - Scripts: call the REST API with the same OAuth token, or with a personal
sm_API key you create under Settings → AI assistants in your profile menu. The OpenAPI document is athttps://app.safemeet.now/api/v1/openapi.json.
The server answers in Code Mode by default: an assistant calls search_tools,
then runs a short program with execute_typescript. Every operation is also a
named tool at https://app.safemeet.now/api/mcp?tools=directory. The full tool
list, with the permission each one needs, is at
https://app.safemeet.now/agent-tools.json.
Signing in and permissions
The assistant sends you to SafeMeet to sign in. You then pick what it may do: read, write, delete, account, support, sign-in connections, API keys and billing. Every permission is checked by default, and each one says what it allows. Untick any you do not want. You can disconnect an assistant or revoke a key at any time from Settings → AI assistants in your profile menu.
Signing in with an assistant follows the same age rule as the app: SafeMeet is for people 13 and older.
What the other person sees
A message your assistant sends carries the label Sent with an AI assistant, so the person you are chatting with knows. See Bots, disclosed for how this sits next to the AI companion label.
Actions that need your say-so
Deleting a message, reporting or blocking someone, filing an appeal, deleting your account, unlinking a sign-in method, revoking a key or a connected app, and cancelling a subscription all come back first with a confirmation token. The assistant has to tell you what will happen and repeat the call with that token. The panic action never asks.
Some answers are yours alone: your birthday and the 18+ confirmation, the reason for a report, whether to keep chatting, and the words of an appeal. The tools ask the assistant to get those from you, never to choose them.
The same checks as the app
A tool call goes through the same permission, sanction, rate-limit and moderation checks as the app. A refused call returns one error shape with a code and, when there is one, the next tool to call.