Comparing LMS MCP servers head to head sounds straightforward. It isn’t possible yet, and the reason why is the most useful thing a buyer can know about this category.
Three names come up most often in this space: Mini Course Generator, Mindsmith and Docebo. Only one of the three has a public page describing what its MCP server does.
That’s not an accusation of anything. Products get announced before they ship, pages move, and capability sometimes lives behind a sales conversation. But it does change the question a buyer should be asking, from which of these is best to how do I tell what any of them actually does.
Try Mini Course Generator free – connect it and find out in one prompt.
Key Takeaways
– LMS MCP servers fall into three classes: knowledge sources (read only), component generators (produce text), and authoring and publishing servers (create the course object). Only the third closes the loop.
– The class determines everything. A knowledge-source server can’t build a course no matter how good the model driving it is.
– Ask for the tool list. An MCP server exposes named actions, and that list answers the capability question definitively where marketing copy doesn’t.
– Verify capability claims against a vendor-published page with a date on it. In this category, a capability description from six months ago is fiction.
– Of the three most-named LMS MCP servers – Mini Course Generator, Mindsmith and Docebo – only one publishes a description of what its server does.
The Three Classes of LMS MCP Server
Almost every question about what one of these can do resolves to which class it’s in.
Knowledge source. The agent can read the system – course catalogues, learner records, content. It cannot change anything. This is the most common class, it’s genuinely useful for questions like “what training has this team completed”, and it produces nothing.
Component generator. The agent can generate a piece of content: quiz questions, a lesson plan, an outline. The output is text, described as suitable for import. You still assemble the course.
Authoring and publishing. The agent creates the actual course object in the platform – lessons, assessments, structure – and publishes it. Some also expose analytics, which is what allows the agent to act on results afterwards.
| Knowledge source | Component generator | Authoring and publishing | |
|---|---|---|---|
| Agent can read the system | Yes | No | Yes |
| Agent can produce content | No | Yes, as text | Yes, as a course |
| Ends with a published course | No | No | Yes |
| Agent can act on learner results | No | No | Only if analytics are exposed |
| You still assemble by hand | N/A | Yes | No |
The distinction that matters commercially is between the second and third rows. A component generator moves your work from writing to formatting. It doesn’t remove it.
Ask for the Tool List
The single most useful evaluation move, and almost nobody does it.
An MCP server exposes named actions. That list is not a marketing artefact – it’s the technical contract between the server and any client, and it’s definitive. If publish_course isn’t in the list, the server cannot publish a course, whatever the page says.
Two ways to get it:
Connect it and ask. Once a server is added, ask the assistant directly: “What tools does this connector expose? List them with what each one does.” It’ll enumerate them, because it has to know in order to use them.
Ask the vendor. “Can you send me the list of tools your MCP server exposes?” is a question with one correct answer and no room for interpretation. A vendor that can’t produce it either doesn’t have a shipping server or has one thinner than the page suggests.
Do this before the demo, not after. A demo is a curated path through the tools that work.
## What Each Vendor Publishes
Vendor documentation was checked in August 2026. In this category, re-check before you sign.
Mini Course Generator – documented. The MCP server is publicly described with a connection endpoint, OAuth authorization, and setup guides for ChatGPT Developer Mode, Claude Connectors and Manus. Class: authoring and publishing, with analytics exposed – the agent creates courses, publishes them, and can read learner results afterwards. This is our own product, and you should weigh the description accordingly and check the tool list yourself.
**Mindsmith** – not documented. Mindsmith does not publish a page describing an MCP server, and its blog does not mention MCP, the Model Context Protocol, or connecting external AI assistants. That doesn’t mean no MCP capability exists, only that there is no vendor-published description of one. Ask them directly and request the tool list.
**Docebo** – not documented. Docebo does not publish a page describing an MCP server. It has publicly discussed agent capability in its platform direction, which is a different claim from a shipping MCP server that external assistants can connect to. Ask them directly and request the tool list.
Udemy Business – a business MCP server was reported as generally available in March 2026, positioned as a knowledge source rather than an authoring surface. Re-confirm current status before relying on it.
If you’re evaluating any of these, the tool-list question resolves in a day what a week of reading pages won’t.
The Questions That Separate Them
Five, in the order that eliminates fastest.
1. Can the agent publish, or only draft?
The class question, and the one that decides most purchases. Ask to see a live URL created during the call, not a preview.
2. Can the agent read learner results?
Plenty of servers now let an agent create a course. Far fewer let it read what happened afterwards, and a write-only connection can’t do any post-publish work. That whole category of value is covered in what an agent does after publish.
3. Whose permissions does it use?
The answer should be yours, through an authorization step you performed. Anything vaguer deserves a follow-up, especially if the server acts across a whole workspace.
4. Which clients are supported?
Remote MCP servers work with clients that support them – Claude, ChatGPT in Developer Mode, Manus, Cursor among others. If your organisation standardised on something else, capability is academic.
5. Can the content leave?
SCORM and PDF export, confirmed before you commit. A server that builds courses quickly into a platform you can’t export from has moved your problem rather than solved it.
Why Capability Claims Go Stale Here
Worth understanding, because it changes how you should read anything written about this category, including this.
MCP is young. Servers ship, get extended, get deprecated and change their tool lists on a scale of months rather than years. A capability list published in March describes March.
Three habits handle it:
Date every claim you rely on. “As of [date], the server exposed these tools.” Without a date, a capability list is a rumour.
Prefer the tool list to the page. The page is written by marketing on a quarterly cycle. The tool list is generated by the server and is true right now.
Re-verify before the contract, not before the demo. The gap between evaluation and signature is often two months, which in this category is long enough for the answer to change.
Our own review of [MCP servers for education and L&D](/blog/best-mcp-servers-for-education/) is subject to the same rule, so re-check anything in it before you sign.
What This Means If You’re Buying
A short version of the process that works.
Start from the class you need. If you want an agent to answer questions about training records, a knowledge source is correct and cheaper. If you want it to build courses, only the third class qualifies, and most of the market isn’t in it.
Get the tool list before the demo. It eliminates candidates faster than any call.
Test with your own material. Every one of these looks good on the vendor’s sample document. The useful test is your actual runbook, which is messier than any demo asset and is what the thing will actually be pointed at.
Check the exit before the entrance. Export format, and whether the content survives you leaving.
If it helps to see what the connected version feels like before evaluating anyone, the fastest route is to connect one and point it at a real document. The plain-English explainer covers what to expect.
The Short Version
A head-to-head comparison of the three most-named LMS MCP servers isn’t possible yet, because two of the three vendors publish no documentation of the capability. That’s worth knowing on its own: this category has more announcement than shipping product.
What replaces it is a framework. Sort any candidate into knowledge source, component generator, or authoring and publishing – only the third produces a course. Then ask for the tool list, which is the technical contract and settles the capability question in a way marketing copy can’t.
Five questions eliminate fast: can it publish, can it read results, whose permissions, which clients, and can the content leave.
And date everything. In a category this young, an undated capability claim is a rumour with formatting. If you want to check ours, the MCP server connects in about a minute and the tool list is one question away.
Frequently Asked Questions
What is an LMS MCP server?
A connection point that lets an AI assistant work with a learning platform. What it can do depends on its class: read the system only, generate content as text, or create and publish actual courses. The three are frequently described in similar language.
How do I find out what an MCP server can actually do?
Ask for its tool list. An MCP server exposes named actions, and that list is the technical contract rather than a marketing description. Connect it and ask the assistant to enumerate the tools, or ask the vendor directly – it’s a question with one correct answer.
Does Docebo have an MCP server?
Docebo does not publish a page describing one. It has discussed agent capability in its platform direction, which is a different claim. Ask them directly and request the tool list.
Does Mindsmith have an MCP server?
Mindsmith does not publish a description of one, on its site or on its blog. That doesn’t prove nothing exists, only that there is no public description of it. Ask them directly and request the tool list.
What is the difference between a knowledge source and an authoring MCP server?
A knowledge source lets an agent read the platform – useful for questions about records and catalogues – and cannot change anything. An authoring server lets the agent create the course object and publish it. Only the second removes the manual assembly work.
How often should I re-check an MCP server’s capabilities?
Before signing, not just before the demo. Servers in this category change their tool lists on a monthly scale, and evaluation-to-contract commonly takes two months, which is long enough for the answer to change in either direction.
