Connector (AI connector)

Anthropic calls them connectors; OpenAI calls them custom connectors behind developer mode. Both are remote MCP servers. You authorise one once with OAuth, and from then on the assistant can list the actions that server offers and call them mid-conversation, acting with the permissions you already hold in that tool rather than a copy of your password.

A vendor directory is a trust signal rather than a boundary. Any remote MCP server can be added as a custom connector by URL whether or not it is listed. Anthropic makes custom connectors available on Free, Pro, Max, Team and Enterprise, with Free limited to one; on Team and Enterprise an Owner adds the connector for the organisation and members then connect their own account.

The security model is worth understanding before the convenience. The model reads tool descriptions as guidance, so a hostile server can place instructions in them — Anthropic’s own documentation warns that “malicious MCP servers may include hidden instructions” and advises connecting only to trusted servers. Treat adding one like installing an app with account access.

The practical limit is not the plan cap but attention. Every connected server adds its tool descriptions to every turn; past a few dozen tools, model accuracy falls and latency rises. Three or four well-chosen connectors outperform twelve.

Meta’s Muse, which opened its connector platform in September 2026, is the odd one out and worth knowing about because the difference is procedural rather than technical. Its Settings › Connectors screen lists only Meta’s built-in connectors and takes no URL: you ask Muse in conversation to build a Custom Connector and hand it a hosted URL, and it builds, tests and saves the integration itself as a reusable skill. The endpoint it consumes is the same remote MCP server every other client takes.