Connectors & plugins: bring your tools to the assistant with MCP

What a connector is

A connector is a bridge between the assistant and a tool you already use, built on MCP — the open Model Context Protocol. Connect one and its actions become tools the assistant can call in every mode: read a Linear issue in Chat, update a Notion page from Cowork, look up a Stripe customer while planning in Code.

Because MCP is an open standard, you’re not limited to what ships in the box: any remote MCP server can be added in a couple of clicks.

Browse the catalog

Settings → Connectors opens a browsable directory of over 160 verified remote MCP servers across nineteen categories — software development, project management, CRM, payments, productivity, design, databases, search and more, each with its real logo so you can find tools at a glance.

Connecting is one click plus the sign-in the service itself asks for: OAuth where the provider supports it, an API key where it doesn’t, and nothing at all for open servers.

Local plugins: .mcpb files

Not everything should be a remote service. Plugins are local MCP servers installed from a .mcpb file — the same format as Claude Desktop extensions — so the ecosystem built for Claude Desktop works here too. A plugin adds tools that run right on your machine, with no server setup.

Drop the file in, review what the plugin provides, and its tools join the same permission system as everything else.

Permissions you actually control

Every connector tool has a per-tool setting: Always allow, Needs approval, or Off. The defaults are sensible — actions that only read run freely, actions that change data ask first — and you can tighten or relax each one individually.

Scheduled routines are stricter still: each routine gets an explicit list of connectors it may use, so an unattended run can never reach a tool you didn’t hand it.

The security model

Connector tokens are stored encrypted on your machine and never leave it — calls go from your computer to the service, with no relay in the middle. And content a connector fetches from the web is treated as untrusted: instructions hidden in a fetched page are not followed as commands, which is the guard against prompt-injection.

Ideas to steal

A few patterns that pay off quickly:

• Triage: read the newest issues from your tracker and draft responses in one pass.
• Docs that stay current: after a work session, have the assistant update the Notion or Confluence page.
• Support digest: search your helpdesk inbox for a topic and summarize what customers actually said.
• Payments lookup: pull up a customer’s subscription state without leaving the conversation.

Frequently asked questions

What is MCP?

The Model Context Protocol — an open standard that lets AI apps talk to external tools. More AI uses it for both remote connectors and local plugins, so anything that speaks MCP can become a tool the assistant may call.

Can I use Claude Desktop extensions in More AI?

Yes — plugins install from .mcpb files, the same format Claude Desktop uses for its extensions, so that ecosystem carries over. A plugin runs locally on your machine with no server setup.

Do connectors get access to my data?

Calls go directly from your computer to the service you connected; tokens are stored encrypted locally and never leave your machine. Read actions run freely, write actions ask first, and you can turn any tool off entirely.

Can scheduled routines use connectors?

Yes, but only the ones you explicitly allow per routine — an unattended run can never reach a connector you didn’t hand it.

My tool isn’t in the catalog — can I still connect it?

Yes. Add any remote MCP server by its URL in a couple of clicks, or install a local plugin from a .mcpb file. The catalog is a curated starting point, not a limit.