AI
Assistants and MCP
Put workflows and app actions in a toolbox, and let Claude, ChatGPT, Cursor or Claude Code call them.
The Model Context Protocol (MCP) lets an AI assistant call tools. VisualAutomate gives it tools to call: your workflows, actions in the apps you have connected, and read-only lookups, gathered in a toolbox.
Toolboxes
A toolbox is a named set of tools with its own address and its own keys. An assistant connected to one sees what is in it, these workflows and those actions, and nothing else on the account. Make one toolbox per assistant or per purpose, and put in it only what that assistant should be able to do.
Toolboxes are managed under MCP in the dashboard.
What a toolbox can hold
- Workflows. The assistant calls one with the arguments the conversation gives it, and gets the result back.
- App actions. Single actions in the apps you have connected, such as sending a mail or creating a contact.
- Built-in reads. Your workflows and how one is built, recent runs, your credits and your connected apps. Never the values typed into a step, and never a token.
Connecting an assistant
In Claude, open Settings → Connectors → Add custom connector. Paste the sign-in address from the toolbox, sign in to VisualAutomate and choose the toolbox on the consent screen. No key is needed.
You can also paste the toolbox's own URL, choose No sign-in, and add an Authorization request header with the value Bearer <toolbox key>.
The toolbox page shows the exact address and key for each client, ready to copy.
Some clients can neither sign in nor send headers, and for those the key can go in the address. An address ends up in logs, history and config files, so prefer signing in or a header wherever you can.
Every call is a real run
A call from an assistant lands in your run list like any other run, with a log, and the credits settle on what it actually cost. An assistant working in your accounts leaves the same trace you would.
Disconnect it, or delete the key, and it stops at the next call.
MCP on the canvas
It works the other way too. Two steps let a workflow use any MCP server:
- MCP: list tools returns the tools a server offers.
- MCP: call tool calls one, with arguments from earlier steps.
Was this page helpful?