Overview
The Model Context Protocol has two directions. In one, an external AI client connects into Taskade. In the other, Taskade is the client and calls out to someone else's MCP server. This article explains the second direction, and where your agents fit inside it.
TL;DR: The outbound direction runs through your automations. The MCP Client connector points an automation at any remote MCP server and calls its tools as steps, on every plan. An agent still cannot call a server on its own, so hand that work to an automation. The inbound direction lives in MCP Server. Build a custom agent →.
How the client direction works
MCP is a shared language for describing tools. A server publishes a list of tools it can run, and a client asks for them by name. When an AI product is the client, the server's tools sit alongside whatever tools that client already has, and the model picks between them. Nothing about the protocol is specific to Taskade, which is why the same endpoint can serve many different AI products.
In a connector model, the server is registered once at the workspace level rather than per agent, so sign-in and tool listing are handled in one place. That is exactly how Taskade ships it. The MCP Client connector holds the server address and its token, and any automation in the workspace can call that server's tools.
What this looks like today
The outbound direction runs through automations, not through the agent itself. You connect the server once, then add steps.
- Open Automations and create an automation.
- Add a step and pick MCP Client.
- Paste the server address, plus a token if the server needs one.
- Add List Tools to see what the server offers, then Call Tool to run one.
An agent can still be part of the picture. Let it read a project, make the judgment call, and pass its answer into the Call Tool step. What an agent cannot do is reach a remote MCP server on its own, without an automation carrying the call.
Full setup, the result fields, and the address rules are in MCP Client Integration.
Why the pattern is useful
The appeal of the client direction is that it saves you from moving data by hand. A research automation can read a wiki where it already lives. A triage automation can file an issue in the tracker your team already uses. An automation can pull a fact from one server, hand it to an agent to reason on, then write that answer to another server. None of that requires exporting anything, because the automation calls the tools where they already run.
The three MCP pages explained
| Page | Direction | What it covers |
|---|---|---|
| MCP Server | Inbound | Claude Desktop, Cursor, Windsurf, and VS Code connecting into your workspace |
| This page | Outbound | How the client direction works, and where agents fit |
| MCP Client Integration | Outbound | Step-by-step setup for the connector, its actions, and its result fields |
The outbound connector works on every plan, with your plan setting how many automation runs you get. The hosted MCP server behind the inbound direction needs a paid plan. The full matrix lives at Pricing.
For agent work alongside this, Agent Tools covers the built-in tools available now, and Integrations covers the 100+ bidirectional connections that move data between Taskade and other services.
Frequently Asked Questions
Can Taskade connect out to external MCP servers?
Yes, through automations. The MCP Client connector holds the server address and calls its tools as steps. An agent cannot reach a server on its own, so let it hand the work to an automation step instead.
What is the difference between the two MCP directions?
Inbound means an outside AI client, such as Claude Desktop or Cursor, connects into Taskade and works with your workspace. Outbound means Taskade is the client, calling a tool that someone else hosts. Both work today, inbound through the hosted MCP server and outbound through the MCP Client connector in automations.
Which plan do I need to use MCP?
The MCP Client connector works on every plan, and your plan sets how many automation runs you get. The hosted MCP server, which lets other AI clients connect into Taskade, needs a paid plan. See Pricing for the full plan matrix.
How can other AI tools connect into Taskade?
That is the inbound direction, where clients like Claude Desktop, Cursor, Windsurf, and Zed connect to Taskade through the hosted endpoint with a point-and-click sign-in and no API key. See Taskade as an MCP Server for the setup steps and plan details.
How can my agents work with outside tools today?
Use the built-in agent tools, automations, and Taskade's integrations. Agent Tools covers what an agent can call right now, including web search, code, and file analysis. For an MCP server or any of the 100+ integrations, put the call in an automation the agent feeds.
Related guides
- MCP Client Integration. Setup steps for calling a remote MCP server from an automation.
- Taskade as an MCP Server. The inbound direction, on every paid plan.
- Custom AI Agents. Build an agent and choose which tools it can reach for.
- Agent Tools. All the tool types an agent can call today.
- Agent Knowledge & Memory. Give an agent standing context to work from.
- AI Models & Credit Costs. Live cost on every tool call.