MCP connections

Custom tools let the AI reach your systems. MCP connections let it reach the visitor's own accounts on other services — with their permission, granted in the conversation.

How it works#

1. You add an MCP server in the dashboard: a name and an https:// URL.
2. A visitor asks something that needs their account there.
3. The AI offers to connect it. The visitor authorises through that service's own OAuth screen.
4. Once authorised, the remote tools become available in that conversation, named mcp_<connection>_<tool>.

The visitor authorises on the third-party service directly. You never see their credentials, and neither does AskVoro — only a token, encrypted at rest.

Read-only, for now#

Only read-only tools from an MCP server are exposed. Anything that would write, delete or otherwise change state is filtered out, even if the server offers it.

That's a deliberate limit for this phase. A support assistant that can read a visitor's data is useful; one that can silently change it is a much larger question, and we'd rather ship the first one properly.

When a grant expires#

Tokens expire and people revoke access. When that happens the AI offers to reconnect rather than failing the conversation — the visitor sees a prompt to reauthorise, not an error.

Adding a connection#

In the dashboard, add the server URL. It must:

• use https://
• contain no credentials in the URL
• contain no fragment

Each connection needs a unique name and URL within your workspace. Names become part of every tool name the AI sees, so keep them short and recognisable — calendar reads better than acme_calendar_integration_v2.

Connections can be disabled without deleting them, which is the fastest way to take one out of service while you investigate something.

When this is worth it#

MCP is the right tool when the data lives in the visitor's account on a service you don't control. If the data is in your database, a custom tool is simpler, faster and doesn't require the visitor to authorise anything.