> ## Documentation Index
> Fetch the complete documentation index at: https://docs.markifact.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Agent vs MCP vs Workflow
> Markifact gives you three ways to run the same marketing operations. This guide explains how they differ and helps you pick the right one for the job.
Markifact runs on a single engine of 500+ marketing operations, the same building blocks for pulling a Google Ads report, creating a Meta campaign, or writing to a Sheet. What changes is **how you reach those operations**.
There are three surfaces:
* **Workflows** run operations as a fixed, visual sequence.
* **Agents** run operations conversationally, letting AI decide the steps.
* **MCP** exposes those same operations to an outside AI client like Claude or ChatGPT.
The operations are identical. You are only choosing who drives them: a canvas you built, an AI inside Markifact, or an AI in a tool you already use.
One mental model: **Workflows are the recipe, Agents are the chef, MCP is the takeout window.** Same kitchen, same ingredients, different way of ordering.
### Watch: which one should marketers use?
***
## The 10-second answer
* Same task, every week, no thinking required? **Workflow.**
* One-off or fuzzy task where you'd rather describe it than build it? **Agent.**
* You live in Claude, ChatGPT, or Cursor and want Markifact's data there? **MCP.**
***
## Side by side
| | **Workflow** | **Agent** | **MCP** |
| ------------------------- | ----------------------------------------------- | ------------------------------------------------------------------------------------------ | ----------------------------------------------------------------- |
| **What it is** | A fixed sequence of steps you build on a canvas | A full marketing assistant inside Markifact that picks and runs the steps | A bridge that lets an external AI client use Markifact operations |
| **Who decides the steps** | You, in advance | The AI, at runtime | The external AI client, at runtime |
| **Where you work** | Markifact canvas | Markifact chat, Slack, WhatsApp, or API | Your own AI tool (Claude, ChatGPT, Cursor, etc.) |
| **Best for** | Repeatable, scheduled automation | Any marketing task, on demand or scheduled, with AI driving the steps | Using Markifact without leaving your AI tool |
| **Runs how** | On a trigger, schedule, or manual run | You send a message, on demand or scheduled | The AI client calls operations mid-conversation |
| **Adapts to input** | No, it runs the same path every time | Yes, it reasons about each request | Yes, the client reasons about each request |
| **Setup effort** | Build once, then it runs forever | Start a conversation | Create a server, connect your client once |
| **Approval model** | Runs automatically once triggered | Human-in-the-loop before changes go live | Client-controlled (approve writes per your client's settings) |
| **Credits** | Per node run (e.g. 1 credit to pull a report) | Token-based on the AI model used, or a flat 1 credit per message if you bring your own key | Per operation run (e.g. 1 credit per report); discovery is free |
***
## Workflows
A **workflow** is an automated sequence you build by chaining [nodes](/core-concepts/node) on a visual canvas. You define every step up front, then it runs the same way every time it's triggered.
**Use a workflow when:**
* The task repeats on a schedule (weekly report, daily sync, monthly rollup).
* You need the exact same steps every run, with no surprises.
* The logic is deterministic: pull this, transform that, send it there.
* You want it to run unattended, with no one in the loop.
**Avoid a workflow when:**
* Every run is different and hard to predict.
* You'd spend more time building the flow than just doing the task once.
**Example:** Every Monday at 9am, pull last week's Google Ads performance, build a Slides deck, and post it to Slack. Build it once, and it runs forever.
Don't want to build it by hand? Describe the outcome and the **Copilot** assembles the workflow by selecting and connecting the right nodes for you.
***
## Agents
An **agent** is a conversational AI inside Markifact, purpose-built for marketing. You describe a task in plain language, and the agent figures out which operations to run, executes them, then surfaces the result for your approval before anything goes live. It is a complete marketing assistant, not just a chat box wired to some tools.
What sets Markifact agents apart:
* **Model agnostic, all in one place.** Run on Claude, GPT, or Gemini, and switch the model per agent without changing how you work. No juggling separate apps for separate models.
* **Bring your own API key.** Plug in your own OpenAI, Anthropic, or Google key and the agent's AI tokens are billed directly through your provider, with Markifact charging a flat 1 credit per message instead of model-based token pricing. See [Bring Your Own Keys](/core-concepts/bring-your-own-keys).
* **Reach it anywhere.** Chat from the **web app**, a whitelabeled **Slack** app, **WhatsApp**, or the **API**, all driving the same operations and connections.
* **Built for marketers, not engineers.** A marketing-friendly UI, results surfaced for review, and a human-in-the-loop approval step before anything publishes.
* **Schedule it.** Run a task on demand or set it to run on a schedule, so an agent can handle recurring marketing work without you starting the conversation each time.
* **Skills, built in.** Package your repeatable playbooks as [skills](/core-concepts/agent-skills) that load on demand, and build new ones with the skill creator. No external setup, it's part of the agent.
* **Multiple agents, shared with your team.** Create as many custom agents as you need, each with its own name, instructions, connected tools, and context (for example, one agent scoped to Client A, another to Client B), then share them with teammates so everyone works from the same setup.
**Use an agent when:**
* The task is one-off or open-ended ("audit this account and tell me what's wrong").
* You'd rather describe the goal than build the steps.
* The right steps depend on what the data shows.
* You want a human-in-the-loop checkpoint before changes publish.
* You want one marketing assistant your whole team can reach from Slack, WhatsApp, or the web.
**Avoid an agent when:**
* You need a strictly identical process to run unattended with zero variation (a workflow is more predictable and cheaper to repeat).
* The task is high-volume and mechanical, where a fixed flow is faster.
**Example:** "Pull last week's Meta Ads performance, find the three worst-performing ad sets, and draft negative-keyword suggestions." The agent decides the steps and asks before it changes anything. Schedule that same prompt to run every Monday and it becomes a standing report, no canvas required.
See [AI Agents](/core-concepts/agent) for models, channels, custom agents, and credits.
### Where agents fit inside workflows
The line blurs in one useful spot: the **[AI Agent node](/nodes/ai/ai_agent)**. It drops an agent into the middle of a workflow, so a mostly-fixed flow can make a dynamic decision at one step. Reach for it when most of your process is deterministic but one step needs judgment.
***
## MCP
**MCP** (Model Context Protocol) is a bridge. Instead of coming to Markifact, you stay in the AI client you already use, Claude, ChatGPT, Gemini, Cursor, or anything that supports MCP, and that client calls Markifact's operations for you during a normal conversation.
**Use MCP when:**
* You already work in an external AI tool and don't want to switch contexts.
* You want to blend Markifact data with whatever else that AI is doing (analysis, writing, code).
* You're a developer or a customer-facing app that needs Markifact operations programmatically per user.
**Avoid MCP when:**
* You want something to run on a schedule with no one present (use a workflow).
* You don't already use an external AI client, in which case the in-app agent is simpler.
**Example:** From inside Claude, "Pull my GA4 sessions for last month, compare them with the month before, and write up what changed." Claude calls Markifact for the data and handles the analysis itself.
See [MCP Servers](/core-concepts/mcp) for setup, permissions, and security.
***
## Agent vs MCP: the one that actually trips people up
Both are conversational and both run the same operations, so they look interchangeable. They aren't. The difference is **where the AI lives and what it comes with.**
* An **agent** is a finished marketing assistant that lives in Markifact. The AI, the channels, the scheduling, the approval step, skills, shareable custom agents, and the marketing-tuned UI are all included.
* **MCP** is a pipe. It gives the operations to an AI client *you* bring (Claude, ChatGPT, Cursor), and that client supplies the AI and the interface.
So the question is rarely "which is more powerful." It is "do I want Markifact to be the assistant, or do I want to plug Markifact into the assistant I already use?"
| | **Agent** | **MCP** |
| ---------------------------- | ---------------------------------------------------------------------------------- | --------------------------------------------------------------- |
| **Where the AI lives** | Inside Markifact | In your own AI client |
| **Who provides the model** | Markifact (Claude, GPT, Gemini, or your own key) | Whatever your client runs |
| **Interface** | Marketing-built UI: web, Slack, WhatsApp, API | Your client's chat window |
| **Scheduling** | Built in, run tasks on a schedule | Not built in, the client must be driven each time |
| **Approval before writes** | Human-in-the-loop, built in | Depends on your client's tool settings |
| **Skills** | Built-in skills and a skill creator for repeatable playbooks | Depends on the client (if it supports skills at all) |
| **Custom, named assistants** | Yes, per purpose or per client, shareable with your team | No, it's one set of tools |
| **Credits** | Token-based on the AI model used, or a flat 1 credit per message with your own key | Per operation run (e.g. 1 credit per report); discovery is free |
| **Best fit** | Marketers who want a ready assistant for the team | Power users and developers who live in an AI tool |
**Choose the agent** when you want a turnkey marketing assistant the whole team can reach, with scheduling and approvals handled for you, and the freedom to pick or bring your own model.
**Choose MCP** when you already work in Claude, ChatGPT, or Cursor all day and just want Markifact's data and actions available there, alongside everything else that tool does.
They aren't either/or. Use the agent inside Markifact for marketing work and scheduled tasks, and add MCP so you can also reach Markifact from your favorite AI tool when you happen to be there.
***
## Pick by scenario
| You want to... | Use |
| ---------------------------------------------------------------------- | ------------------------------------------------------ |
| Send the same weekly report automatically | **Workflow** |
| Run a quick, one-off account audit | **Agent** |
| Bulk-create 50 ads from a spreadsheet on a trigger | **Workflow** |
| Ask "what changed in my campaigns this week?" and get a written answer | **Agent** |
| Work in Claude or ChatGPT and pull Markifact data there | **MCP** |
| Build a customer-facing app that runs operations per user | **MCP** (or the [API](/api-reference/getting-started)) |
| Mostly fixed process with one step that needs AI judgment | **Workflow + [AI Agent node](/nodes/ai/ai_agent)** |
| Let a non-technical teammate trigger a task from Slack | **Agent** (Slack channel) |
***
## They are not mutually exclusive
These surfaces share the same operations and the same [connections](/core-concepts/connections), so they layer naturally:
* A **workflow** can call an **AI Agent node** for a dynamic step.
* An **agent** can do the exploratory work, and once you've nailed the steps, you rebuild it as a **workflow** to run on a schedule.
* **MCP** lets your external AI trigger Markifact while you keep building workflows in the background.
A common path: prototype with an agent, productionize as a workflow, and reach for MCP when you want it all from your own AI tool.
***
## FAQ
The **agent** is a complete marketing assistant that lives in Markifact: it brings its own AI (Claude, GPT, Gemini, or your own key), a marketing-built UI, channels (web, Slack, WhatsApp, API), scheduled tasks, and a built-in approval step. **MCP** is just a pipe that hands the same operations to an AI client you already use, like Claude or Cursor, where that client supplies the model and the interface. Pick the agent when you want Markifact to *be* the assistant; pick MCP when you want to plug Markifact into the assistant you already use. See [Agent vs MCP](#agent-vs-mcp-the-one-that-actually-trips-people-up) above.
Workflows are deterministic, cheaper to run repeatedly, and need no one present. When a process is fixed and recurring, a workflow runs the exact same path every time with no token cost for reasoning. Agents shine when the steps aren't known in advance or change run to run.
No. Workflows, agents, and MCP all use the same [connections](/core-concepts/connections) in your workspace. Connect Google Ads once and every surface can use it.
* **Workflows** charge per node that runs. Pulling a Google Ads report costs about 1 credit; account selection and field listing are free.
* **Agents** are priced on the AI model you use (token-based, with a multiplier per model, so a heavier reasoning model costs more than a lighter one). Bring your own API key and the AI tokens are billed through your provider instead, with Markifact charging a flat 1 credit per message.
* **MCP** charges per operation it runs, at the same rate as the matching workflow node. Pulling a report is about 1 credit; searching for operations, reading inputs, and listing connections are free.
A quick rule: pulling or writing data costs around 1 credit per action, while discovery and setup steps are free. See [Credits & Usage](/core-concepts/credits) for the full tables.
Yes, and it's a good pattern. Explore a task with an agent, then once the steps are clear, rebuild it as a workflow to run on a schedule. Nothing is locked in, because they all sit on the same operations.
***
## Related
Build a fixed, visual automation that runs on a schedule or trigger.
Describe a task and let Markifact's AI pick and run the steps.
Use Markifact operations from Claude, ChatGPT, Cursor, and more.