If you use an AI tool at work, you may have seen the letters MCP in its settings. They stand for Model Context Protocol, an open standard that lets an AI tool use other software: not just talk about it, but look things up in it and make changes to it. For automation, that's a bigger shift than it sounds. This explainer covers what MCP is, what changes when your AI tool can build automations rather than describe them, and what to check before you connect one, using Lodol's own MCP server as the example.
The problem it solves
On its own, an AI model can only produce text. To look up an order, add a row to a sheet or start a process, it needs a way to reach the software that holds them. Without a shared standard, every pairing of an AI tool and an app needs its own custom connector, built and kept working by someone.
MCP replaces all those one-off connectors with a single shared way of plugging in. A piece of software offers an MCP server once, and any AI tool that supports MCP can use it. Claude, ChatGPT, Cursor, VS Code, Claude Code, Codex and many others speak it, which is why one web address connects all of them to Lodol.
How it works, in three parts
- The server belongs to the software being used. It offers a list of tools, each with a name, a plain-words description of what it does and the details it needs. Lodol's lets an AI tool list your workflows and explain them, search the apps and actions Lodol connects to, check a workflow before it's saved, save it, run it, and read how a run went.
- The client is your AI tool. It reads that list, and when your request needs one of the tools, it calls it, reads what comes back and decides what to do next.
- The sign-in is yours. When you connect, you sign in to the software yourself and approve what the AI may do. With Lodol, you sign in with your Lodol account, so there's no key to copy, then choose the workspace and untick any permission you'd rather it didn't have.
From describing an automation to building one
Ask an AI tool without MCP how to automate something, and the best it can do is hand you instructions or a script. Then you're the one who finds somewhere to run the script, sets a schedule, stores the passwords it needs and checks a log nobody reads. The AI did the easy part.
Connected to Lodol, the same request becomes a working workflow. Type "When a Stripe payment comes in, add it to our revenue sheet and post it in #sales", and your AI picks the apps, builds a New Stripe payment trigger, an Append row step for Google Sheets and a message step for Slack in your workspace, and gives you the link. You open it, connect the apps, turn it on, and every payment from then on is logged and announced. Lodol checks what your AI builds against the apps, actions and fields that really exist, and lists anything you still need to fill in.
Why the AI shouldn't be the one running it
Here's the part that matters most. Because MCP lets an AI do things, it's tempting to have the AI do the job itself every time: read each new email, decide, write the row. For a one-off, that's fine. For something that has to happen every day, it causes three problems.
- It can take a different path each time. A model works the job out afresh on every pass, so the hundredth run may not do what the first one did.
- It costs on every run. The AI spends tokens working out the same steps again and again.
- It needs the AI every time. Something has to start the AI for each email, and there's no record of what it did unless you build one.
Building a workflow avoids all three. Your AI works the steps out once, and Lodol runs exactly those steps on every trigger, whether or not your AI tool is open. Your AI only spends tokens while it builds or changes the workflow, and a model is called on a run only at a step that asks for one, such as reading a receipt. Every run is recorded step by step. That's the design we think AI belongs in: intelligence where reading and judgment are needed, and fixed steps everywhere else. Our CEO wrote about why in Agents Are Having a Moment. Reliability Is Having a Reckoning.
What stays out of the chat
Handing an AI tool access to your work deserves some care, so here is what Lodol keeps back:
- Your keys. Apps are connected in Lodol itself, keys are stored encrypted, and workflows use them by name. Your AI builds with the name; the value never appears in the conversation.
- The jobs that belong to people. Connecting apps and managing members, billing and secrets stay in Lodol, done by someone in your workspace.
- Anything beyond your role. Your AI acts as you, and can never do more than your own role allows.
- Deleting. As of today, an AI tool can delete a workflow only if you grant it the Delete workflows permission, which stays off until you turn it on, and it's asked to check with you first.
And nothing your AI does is hidden. Its changes go into the workflow's version history like anyone else's edit, so on paid plans you can see exactly what changed and roll back in a click.
Questions to ask of any MCP connection
Not every MCP server makes the same choices about what an AI may do. Before you connect one, it's worth asking:
| Ask | Lodol's answer |
|---|---|
| What can it do, and can I switch parts off? | Find, explain, build, change and run workflows, and read runs. You choose its permissions when you connect. |
| Does it act as me, or as something bigger? | As you, within your role. |
| Where do my passwords and keys go? | Into Lodol, encrypted. The AI only sees their names. |
| Is there a record, and can I undo it? | Every change goes into the version history, and every run is recorded step by step. |
| Does anything destructive need my say-so? | Deleting is off unless you grant it, and your AI asks first. |
| Once it's built, does the work still need the AI? | No. Lodol runs the workflow on its own. |
Try it
Connect AI, under Workflows in Lodol's sidebar, has the one step for each tool: a click for Claude, Cursor and VS Code, a single command for Claude Code and Codex, and one web address, https://mcp.lodol.com/mcp, for ChatGPT and anything else that supports MCP. Then ask for something small, like a message in Slack whenever a particular customer emails. Our guide to asking your AI for a workflow covers how to brief it so the first version is close to right.
Connect your AI in a minute.
Pick your tool on our homepage, or follow the setup steps in the docs. See what it looks like when your AI builds a workflow, or start free.