Every automation that thinks has to run on some model. Taskade gives you three ways to choose one, and they are not a ranking — they answer different questions.
Short answer: use a Taskade AI step by default, name a provider when the job depends on that specific model, and use the MCP Client connector when the flow has to reach a system no connector covers yet.
What are the three ways to run a model in a workflow?
| Approach | What it is | Right when |
|---|---|---|
| A Taskade AI step | Ask AI, Generate with AI, Categorize with AI, or a named AI agent. No key, no vendor account, no billing to set up. | Most of the time. You want the work done, and you would rather not own the model decision. |
| A named provider action | An OpenAI, Anthropic Claude, Google Gemini, Groq, or OpenRouter step you add to the flow and point at your own provider account. | The model identity genuinely changes the outcome — see the list below. |
| The MCP Client connector | Two actions, List Tools and Call Tool, that let a flow talk to an MCP server you point it at. | The flow needs a system Taskade has no connector for, and that system already exposes an MCP server. |
All three can sit in the same flow. Several of the 16 automations below do exactly that: a fast model screens, a frontier model handles what it flags, and an agent writes the result where a person will read it.
Which model does a Taskade AI step use?
Taskade runs on 15+ frontier models from OpenAI, Anthropic, Google, and open-weight providers. The default is Auto, which picks a model for the job so you do not have to, and you can set a specific model on an agent whenever you want one. Either way the step works the moment you add it — there is no key to paste and no provider account to open.
This is the path to start on. Build the flow, see whether the answers are good, and only then decide whether the model deserves a decision of its own.
When should you name a provider instead?
Name one when the answer to does it matter which model does this? is genuinely yes:
- An internal policy names it. Some teams may only send a class of document to one approved provider. Putting that provider in the step makes the choice visible in the flow instead of implied by a default.
- The input is very long. A contract, a specification, a research report. A model chosen for long-document reading holds the whole thing at once rather than in fragments.
- The output has to match one exact shape. OpenAI, Anthropic Claude, Google Gemini, Groq, and OpenRouter each ship an Extract Structured Data action, which returns fields in the schema you define rather than prose you then have to parse.
- Volume makes cost the deciding factor. Send the high-volume first pass to a fast, inexpensive model and reserve a frontier model for the cases that earn it. A branch is how you say that in one flow — see automation logic.
These are options Taskade supports side by side. Each of those providers is good at something, which is the entire reason you get to name one; nothing on this page asks you to rank them.
What does the MCP Client connector do?
MCP, the Model Context Protocol, describes a connection between an AI system and a tool server. The direction matters, and Taskade has more than one — so read the arrow, not just the three letters.
- A flow calling out to an MCP server. That is the MCP Client connector, and it is what every MCP automation on this page uses. It ships two actions, List Tools and Call Tool, and it is available on every plan, including Free. It is an automation step: it belongs to the flow, alongside your other actions, and you point it at the server yourself.
- An outside client calling in to Taskade. Taskade also hosts its own MCP server, so a compatible client can read and write your workspace from outside. That is the opposite arrow and a paid-plan feature — see pricing.
The practical use for the outbound direction is reach. Your warehouse system, your internal ticketing tool, your in-house database — if it has no connector yet but exposes an MCP server, a flow can list what that server offers and call one of its tools as a step, with the result flowing into the next step like any other.
Who should use these automations?
Anyone who has already built a flow and is now asking what runs the thinking part of it. Finance and operations teams standardizing extraction. Support and community teams with more volume than a frontier model is worth spending on every message. Engineering teams reviewing more than they can read. And teams whose system of record has no connector yet, who need MCP to close the gap today rather than wait for one.
You do not need to know anything about models to start. Add a Taskade AI step, watch it work, and come back to this page the day one of the four reasons above applies to you.
How do you set one up?
- Open the Automations tab in your workspace and add a trigger — a new email, a form response, a completed task, a schedule.
- Add the thinking step: a Taskade AI step, a named provider action, or an MCP Client call.
- If you named a provider, connect that provider account on the step the first time you add it; a Taskade AI step needs nothing.
- Send the result somewhere a person will see it — a task, a comment, a channel, a draft reply.
- Or describe the outcome at Taskade Genesis and let it wire the trigger, the steps, and the agent for you.
Where should you go next?
- All automation categories — the full index, and automations for the hub.
- AI agent automation — when the thinking step should be a named agent that keeps knowledge and memory between runs.
- Automation logic and automation triggers — the branches, loops, filters, and events these flows are built from.
- Taskade connects to 100+ apps in both directions: triggers pull events in, actions push data out.
