Your data

MCP tool

MCP tool calls a tool on an MCP server you connect: one of your own systems (pricing, stock, customer records) or an AI agent that speaks the Model Context Protocol. Call it once for each item, with arguments filled from the item’s fields, or once for the whole run, and the answer is saved on every item for the blocks after it. It calls only on real runs.

In the editor: Calls a tool on an MCP server you connect, for example your own system or an AI agent, and adds its answer to each item.

Settings

MCP server

An account you connect

An MCP server connection from the Connections page: its address and, if it needs one, a token (kept sealed). Testing the connection lists the tools it offers.

Tool

Text

The tool’s name on the server. Testing the connection lists the tools it offers.

Needed before the flow can run.

Call it

Pick one

Once for each item (each gets its own answer), or once for the whole run (every item gets the same answer).

  • Once per item A call for each item, up to 100 a run, 4 at a time.
  • Once per run One call for everything; the arguments can use {{count}} and {{items}} (the first 100 items).

Arguments (JSON)

Text, several lines

Use {{field}} to fill in a value from each item, e.g. {{name}}, {{lat}} or {{lng}}.

Needed before the flow can run.

Save the answer as

Text

The name of the field the answer is saved in. The blocks after this one get it.

Needed before the flow can run.

This tool changes things

On or off

Off: the tool only looks things up, so it is called in its turn, in test runs too, and its real answer goes on to the next blocks. Switch on for a tool that writes, books or sends something.

What comes out

Each item that leaves the block carries these fields, to use in the blocks after it as {{field}}.

Each item

With Save the answer as left as tool_result. Every item keeps its own fields too.

  • tool_result: The tool’s answer: its text, or its structured answer as JSON text, up to 4,000 characters.
  • tool_result (on an error): Once per item: "Error: " and what the tool said, on that item only; the other calls carry on.

How to use it

Retail

The delivery fee for every new store

A retailer prices deliveries with its own service and wants the fee for each new store that opens in its area.

The arguments {"name": "{{name}}", "lat": "{{lat}}", "lng": "{{lng}}"} send each store’s own details; the fee comes back as a field on the store.

B2B sales

An agent drafts the first message

A sales team runs its own AI agent behind an MCP server and wants a first message drafted for each new lead.

Before: Reps wrote every first message from scratch.

With the flow: Each new lead arrives in Slack with a draft the rep can edit and send.

FMCG distribution

Stock levels at the nearest warehouse

Each account carries the stock answer from the company’s own system, so planners see shortages by customer.

Banking and payments

Is this merchant already a customer?

A bank exposes a lookup_customer tool over MCP instead of opening its core database.

The bank decides what its tool answers; the flow only sees the answer, not the database.

Logistics

One summary call for the whole run

Once per run sends every closure in one call; the planning tool answers once and the answer rides on every item.

Healthcare

Check each clinic against the provider registry

A health insurer wraps its provider registry in an MCP server.

Before: Analysts checked new clinics against the registry by hand.

With the flow: Each new clinic comes back with the registry’s answer, and the clinics it does not know are flagged with an error note.

Good to know

  • Only real runs call the tool: Run once, a schedule, a webhook, new rows or the API with sending on. Test runs and the editor give sample answers.
  • Once per item makes up to 100 calls a run, 4 at a time; once per run makes one call, with the first incoming item’s fields plus {{count}} and {{items}}.
  • In the arguments, a value that is exactly "{{field}}" is sent as that field’s own value (a number stays a number); inside longer text it is filled in as text.
  • An error from the tool is saved on that item and the step says so. Once per run, an error fails the step.
  • The server must be on the public internet; private and local addresses are refused. A personal MCP connection can be used only in flows its owner owns.
  • Answers are kept as text (structured answers as JSON), up to 4,000 characters; use Custom code or xMap LLM after it to pull out parts.
  • Free: calling your own tools costs no credits.

Works well with

  • Map business scan: Finds businesses on online maps live: name, category, rating, address, phone, website and whether it is open.
  • My data: Brings in a table you uploaded or a list you saved on the My data page: your stores, customers, target accounts or sites.
  • New rows: Starts the flow when new rows appear in a Google Sheet or in a table in My data.
  • Webhook / API: Starts the flow when another system sends a POST request with a JSON body to this flow’s webhook URL.
  • Custom code: Your own JavaScript, for what the other blocks cannot do: change, add or drop fields, filter and reshape the items.
  • xMap LLM: Tell the AI what to do in plain words.
  • Slack: Channel-based messaging platform
  • Email: Sends one email per run with all the items, or one email per item.

Try it

Build a flow with MCP tool on the home page, no account needed, or book a call. Try it live