MCP Tool
Call one tool of your own MCP server as a fixed workflow step
The MCP Tool node calls exactly one tool of one of your custom MCP servers, with the values you enter in the node. It works like an integration action: every run makes the same call, with no AI model in between. You find it in the node picker under Integrations, if you have the Use MCP servers permission.
If a model should decide itself whether and which tool to call, give the tool to the AI Agent as a tool instead.
Configuration
- MCP server: one of your connected MCP servers. The node offers the servers you can also pick in the AI Agent.
- Tool: one of the server's tools. If the picker does not know the tools yet, it asks the server; Refresh reloads the list.
- Inputs: the node shows one field per argument of the tool, with the name, description and "required" flag the server declares.
| Tool argument | Field in the node |
|---|---|
| Text | Text field |
| Number | Number field (an integer must be a whole number) |
| Yes/no | Switch |
| Fixed choice | Dropdown |
| Object or list | JSON field |
| Complex description | one single JSON field Arguments for all arguments |
Variables such as {{trigger.outputs.orderNumber}} work in every field, as in any other node. An empty optional field is not sent to the server.
If the server later changes the tool's arguments or outputs, the node shows a notice. Update fields applies the new version and keeps what you already entered where the field still fits. Nothing changes on its own.
Outputs
| Output | Type | Description |
|---|---|---|
result | object | The tool's structured result if it returns one, otherwise its text. Text that is JSON is read as an object |
content | array | The content as the server returns it. Images and files are stored as workflow files and are available under file |
isError | boolean | Whether the tool reported an error; always false on a successful run |
A later node reads a field of the result with {{mcpTool.outputs.result.fieldName}}. If the server describes its tool's outputs, the variable picker shows those fields before the first run.
Errors and approval
If the tool reports an error, the node fails and the run log shows the server's message. With Use error output the workflow takes the error path instead, and the message is available there as {{mcpTool.outputs.error}}.
A selected tool counts as approved; nobody is asked in a workflow. The node fails with a message naming the reason when
- the person who owns the connection set the tool to Never allow under Permissions,
- they lack the Use MCP servers permission or the organization switched the server off for them,
- the server no longer offers the tool under that name.
If the tool changes data, it is best to switch on Requires approval for the node.
With the workflow assistant
The assistant in the workflow builder knows your MCP servers and their tools. Ask it for example "Use my Orders server to create an order for every form submission", and it adds the node and fills in its fields.