Build and Run Your Entire LMS From an AI Assistant

Every LMS has a dashboard. It is where you go to create a course, add a lesson, enroll a cohort, fix a typo on the homepage, publish a post. And it is where a surprising amount of that work quietly goes to die, because it means clicking through tabs, remembering which panel holds which setting, and repeating the same ten-step ritual you did last week.

I am Amir. I have spent thirteen years building learning platforms, and I have watched capable people lose whole afternoons to LMS admin. So we shipped something I have wanted for a long time: Cubite now runs over the Model Context Protocol (MCP), which means you can connect Claude, ChatGPT, or Gemini to your site and just say what you want. "Create a course called Food Safety Basics with three lessons, then enroll these five people." It does it, against your real site, scoped to your permissions.

I am not going to tell you we invented this, or that we are the only LMS that talks to an AI assistant. That would be a lie you could disprove in one search, and a few big platforms got here before us. What I will argue, with the honest competitive map and a plain look at what our tools actually do, is that Cubite lets you operate more of your platform from the assistant than almost anyone else. Not just ask it questions, not just draft, but build and run the whole thing. You decide whether that matters to you.

What an "MCP LMS" actually is

Plain version. MCP (the Model Context Protocol) is an open standard, introduced by Anthropic and now supported across Claude, ChatGPT, and Gemini, that lets an AI assistant call real tools on an outside system. An "MCP server," sometimes called a connector, is the adapter that exposes those tools. Cubite ships one. When you connect it, your assistant gains a set of Cubite tools it can call on your behalf: create a course, enroll a learner, publish a blog post, and so on.

Here is the mental model that helps. The assistant stops being a place you copy answers out of, and becomes a place you get work done. You describe the outcome in the words you actually use. The assistant translates that into the right tool calls. Cubite runs them against your site and reports back what changed. There is no plugin to install on your side and, if you use the browser sign-in, no key to copy or store.

Everything runs against exactly one site (the one your credential is bound to) and only within the permissions you granted. A connection that can read courses but not delete them simply does not have the delete tool. More on that below, because the security model is the part I most wanted to get right.

What you can actually do from the chat box

Here is the honest inventory, not a teaser. Cubite exposes 54 tools today, and your assistant only sees the ones your connection's scopes allow. Grouped by what they touch:

• Courses and course content. Create and update courses (title, dates, price, credits, visibility, prerequisites, instructors, certificate settings), then author the actual lessons: sections, units, rich text, images, and graded assessment blocks. Delete a course, with a guard if learners are already enrolled.
• Learning paths. Sequence several courses into a guided journey with completion rules, prerequisites, paid bundles, and path-level certificates. List the learners on a path with their progress, or enroll someone by email.
• SCORM and xAPI packages. Point a tool at a hosted Storyline, Rise, or iSpring export. Cubite fetches it, validates the manifest, extracts it, and detects whether it is SCORM or xAPI. Drop the returned id into a course and it is playable.
• Blog, pages, and the homepage. Create and publish blog posts (this very post was published through the MCP), edit any page, and rebuild the tenant homepage's hero, features, and call-to-action sections.
• Enrollments and cohorts. Enroll one learner or bulk-enroll up to a hundred, optionally with a welcome sign-in link. Group learners into cohorts with seat caps and bound courses or paths, so adding a member auto-enrolls them.
• Question banks. Build reusable banks and add typed questions (single answer, multiple answer, true/false, fill in the blank, matching, numerical, and more) that your quiz blocks pull from.
• Media and site settings. Upload an image or a file, and update the safe site settings: theme, fonts, logos, SEO metadata, contact details, and layout.

A real example, start to finish. In Claude, connected to a test site, I typed: "Create a course called Kitchen Safety, add three lessons (Handwashing, Cross-Contamination, Cooling), and enroll amir@example.com." Claude called the create-course tool, set the content three times, ran the enroll tool, and came back with the course URL and a confirmation the learner was in. No dashboard, no tab hunting. That is the whole point.

Who else got here first: an honest map

I audited the field, because you deserve the real picture instead of a straw man. Several serious platforms already ship an AI-assistant connector, and I will name them plainly.

• LearnDash ships a first-party MCP server. It does full course, lesson, and topic create/read/update/delete plus enrollment management, with read access to users and quizzes (LearnDash KB).
• Docebo launched AgentHub, an official MCP surface it markets as "your learning, inside the AI your team already uses." Today it leans learner-facing (progress, certificates, library search), with manager and admin write tools rolling toward general availability in mid-2026 (Docebo).
• Kajabi offers an official MCP that reads and writes across pages, emails, offers, courses, and contacts, though the writes land as drafts you approve (Kajabi).
• Thinkific has an official "connect to Claude" integration, currently read-oriented (courses, learners, enrollments, products) and gated to Claude's paid tiers (Thinkific).
• Canvas built IgniteAI, a native agent on top of MCP, and there are rich community Canvas MCP servers besides (Instructure).
• Moodle and Open edX are reachable through a community plugin and a framework example, respectively, rather than a turnkey product. TalentLMS and LearnWorlds are reachable through third-party connectors like Zapier.

So this is not a "nobody does it" story. It is a "the category is moving fast, and the shipped connectors differ a lot in what they let you touch" story. And that difference is where Cubite actually stands apart, stated as scope rather than slogans:

• Most of the connectors above let you query your platform or draft changes. Cubite is write-through: when you publish, it publishes.
• Most cover a slice (courses, or learner queries, or marketing pages). Cubite covers the whole build surface: courses and content, learning paths, question banks, SCORM and xAPI, plus the site itself (pages, blog, homepage, settings). You can stand a site up, not just poke at one.
• Several are single-assistant (Thinkific is Claude-only, on paid tiers). Cubite's server is standard MCP with OAuth 2.1, so Claude, ChatGPT, and Gemini all work, and a scoped key covers CLI and server automations.

If breadth of what you can operate is what you care about, that is the comparison I would make. If you mostly want your assistant to answer "how is this cohort doing," several tools, ours included, can do that.

Why Cubite can do this: it was headless first

Here is the mechanism, because I distrust features that cannot explain themselves. Cubite did not bolt an AI layer onto a monolith. It has been a headless, API-first LMS from the start: every capability lives behind a real, documented public API. The MCP server is a thin, honest translation of that API. When the assistant calls the enroll tool, it hits the same endpoint our own admin uses, with the same validation and the same tenant scoping.

That matters for trust. There is no separate, sketchy "AI backdoor" with its own rules. The assistant is just another authenticated client of an API that already existed, gated by the same permissions. It is also why we could add the connector without re-plumbing the product: the surface was already there.

What it is not, yet (on purpose)

The house rule here is that I would rather you trust me than oversell you, so here is the honest scope.

• It does not pull analytics back. You can read enrollment status and a learner's progress on a path, but there is no tool that returns xAPI statement data, quiz-score reports, or dashboards. "Upload your SCORM from chat" is true. "Ask your assistant for your xAPI analytics" is not, yet.
• Editing is versioned, not live-collaborative. Setting course or page content appends a new version, exactly like the admin editor's save history. It is not a real-time shared cursor.
• It is single-site by design. One connection equals one site. There is no cross-tenant management from a single credential, on purpose.
• It cannot touch secrets. No payment keys, SSO config, analytics IDs, or custom CSS. Those stay in the admin UI, where they belong.

Those are deliberate lines, not shipping gaps I am hiding. When the analytics-read tools are ready, I will say so.

How to connect it in about two minutes

Your site's endpoint is https://your-domain/api/mcp/mcp (your Cubite subdomain, or your custom domain if you have one).

• Claude or ChatGPT: open Settings, then Connectors, then add a custom connector with that URL. It opens a Cubite sign-in, you pick the site, you approve the permissions. Nothing to store.
• Gemini, a CLI, or a server: create a scoped ck_... key in the admin and send it as a bearer header. Keys can be IP-restricted and set to expire.

Either way it is scoped to one site and revocable at any time. And because publishing a post through the MCP also fires IndexNow, content you create this way is discoverable by AI search and assistants almost immediately. That is a pleasant loop: the assistant that helps you write a post also helps the world find it.

See it work on your own courses.

Start a free Cubite trial, connect it to your assistant, and build a course by typing one sentence.

Start a free trialSee how migration works

Where Cubite fits

If you run a serious learning program and you are tired of the dashboard being the slowest part of your day, this is for you. It is the same instinct behind our command palette: stop making people memorize where things live. If you are migrating off a fragile plugin stack and want to stand up courses fast, an assistant that can build them is a genuine accelerant, and it pairs well with how we already help people move off LearnDash. And if you would rather keep clicking, that is fine too. The admin is not going anywhere. Some of us just talk to it now.

I will not claim this replaces every workflow, or that it is the only LMS with an MCP server, because neither is true. What I will say is that we built it because we wanted it, it is live today across the major assistants, and you can try it on your own courses in a couple of minutes. That is the honest version.