Connect

MCP Client Integration

Updated 2026-08-11·7 min read

Overview

The MCP Client connector points a Taskade automation at any remote MCP server and runs that server's tools as steps. Paste the server address once, and every tool it publishes becomes something your automation can call, with the result flowing into the next step like any other action.

TL;DR: Add the MCP Client connector to an automation, paste an MCP server address, and call that server's tools as steps. List Tools shows you what a server offers. Call Tool runs one and hands the result to the next step. Available on every plan. See Pricing for run limits.

This is the outbound direction of MCP, where Taskade reaches out to someone else's server. For the inbound direction, where Claude Desktop or Cursor connects into your workspace, see Taskade MCP Server.

Connect an MCP server

  1. Open Automations in your workspace and click Create.
  2. Add a trigger, then add a step and pick MCP Client.
  3. Click Connect.
  4. Paste the Server URL. It must start with https://.
  5. Paste a Bearer Token if the server needs one. Leave it empty for open servers.
  6. Save. Taskade checks the connection by asking the server for its tool list, so a wrong address or an expired token fails right there instead of at run time.

Your token is masked in the interface and stored encrypted. One connection covers the whole workspace, so every automation that talks to that server reuses it.

Tip: Your own Taskade MCP server lives at https://www.taskade.com/mcp. Point the connector at it to have one automation work with another workspace.

Actions

Action What it does Returns
List Tools Asks the server what it can do tools (name, description, input schema) and count
Call Tool Runs one tool and waits for the result text, content, structuredContent, isError

List Tools

Run List Tools once while you are building. It returns every tool the server publishes, with the name, the description, and the fields each one expects. Use that to decide which tool you want and what arguments to send.

Call Tool

Call Tool has two fields.

  • Tool. A dropdown loaded live from the server, so you pick from what that server actually offers. You can also type a name by hand when the server is temporarily unreachable.
  • Arguments (JSON). A JSON object matching the tool's input schema, for example {"query": "hello"}. Leave it empty for tools that take no arguments.

Reading the result

Field What it holds
text The readable text the tool returned, joined into one string. Use this for messages, summaries, and anything a person reads.
content The raw blocks the tool sent back, in order.
structuredContent Machine-readable data, but only for tools that publish an output schema. It is empty for every other tool, so branch on it only when you know your tool provides one.
isError true when the server flagged the result as a failure. Add a filter on this if you want a fallback path.

Images, audio, and other binary results come back as a link. Taskade uploads the file and puts the address in the result, so a later step can reference the link instead of carrying the file through your automation.

Which servers work

The connector speaks the standard protocol, so a public MCP server works if it is open or accepts a bearer token. Taskade tries the modern transport first and falls back to the older one automatically, and you never pick between them.

A server that expects you to sign in through a browser window, rather than accept a token you paste, cannot be connected here. Check whether the server offers a plain token before you wire it up.

A few addresses are refused on purpose:

  • The address must start with https://. Plain http:// is rejected.
  • Sign-in details cannot be embedded in the address itself. Use the Bearer Token field.
  • Addresses that point at a private network, such as localhost or an internal company address, are rejected. The server has to be reachable on the public internet.

If a server accepts the connection but never finishes the handshake, the step stops after 20 seconds with a message telling you to check the address.

Common patterns

You want to Build this
Pull a fact from another tool mid-automation Trigger, Call Tool, then write the text into a project
Handle a tool that can fail Call Tool, then a filter on isError, then two branches
Feed structured data onward Call Tool, then map fields out of structuredContent
Let an agent decide, then act Ask an agent for the input, then Call Tool with that answer
Audit a new server before wiring it up List Tools on its own, run once, read the output

Frequently Asked Questions

What is the MCP Client connector?

It is an automation connector that lets a Taskade automation call tools on a remote MCP server. You connect the server once with its address and an optional token, then add List Tools or Call Tool steps to any automation in the workspace.

How is this different from the Taskade MCP server?

Direction. The Taskade MCP server lets outside AI tools such as Claude Desktop and Cursor connect into your workspace. The MCP Client connector goes the other way, letting your automations reach out to somebody else's server.

Can my AI agents call an external MCP server directly?

Not directly. Agents work with their large built-in toolkit and your own custom tools. To reach an MCP server, put the Call Tool step in an automation, and let the agent hand its answer to that step. See Taskade Agents and External MCP Servers for the full picture.

Why is my server URL rejected?

The address must start with https:// and point at a server on the public internet. Private addresses such as localhost or an internal company hostname are refused, and sign-in details cannot be embedded in the address. Put any token in the Bearer Token field instead.

Why is structuredContent empty?

Only tools that publish an output schema return structuredContent. Everything else returns its result in text and content. Run List Tools to see what a tool declares before you build a branch on it.

Which plan includes the MCP Client connector?

Every plan. There is no upgrade badge on these steps. What changes between plans is how many automation runs you get, so a busy automation is the thing that wants a bigger plan, not the connector itself. See Pricing for the full matrix.

Was this helpful?