Moodle plugin · Teams and Copilot · ChatGPT · Claude
Moodle is the most widely used distance-learning platform: universities, schools, healthcare organisations and training providers all run on it.
MCPCA connects Moodle to AI assistants: from existing material or from scratch, describe a page, a quiz or a whole course, typed or spoken, and find it ready on the platform, without Moodle's editor and with nothing to install. Students, in turn, query their own courses for clarification and revision.
Everything inside your organisation's own infrastructure
Everyone signs in with their own institutional account and acts only within the permissions Moodle grants them.
Turn these notes into a page in section 3 of Clinical Nursing, with a manual completion tick.
get_my_coursescreate_page I laid the text out like this:
Yes, but keep it hidden until I have reviewed it.
In brief
AI assistants such as Microsoft Copilot, ChatGPT and Claude can hold a conversation, but they know nothing about the content of your organisation's courses and cannot act on it. MCPCA is the missing link.
Moodle is the open source platform behind a large share of online training, in Italy and worldwide: courses, materials, assessments and activity tracking. MCPCA installs inside it like any other component: no extra servers, no external services.
It uses MCP (Model Context Protocol), an open standard adopted by AI assistants and agents such as Microsoft Copilot, ChatGPT and Claude. The assistant can therefore read and write in the courses published on Moodle for which the signed-in user holds the relevant permissions, and only with that user's explicit approval.
Each person signs in with their own institutional account. The plugin performs only the operations that person could perform through Moodle's web interface: a teacher creates and edits in the courses they run, a learner consults the ones they are enrolled in.
Where it fits
The plugin knows nothing about your sector: it knows Moodle. MCPCA grew out of a university teaching need, but the same tools apply wherever teaching content has to be produced and maintained online.
Degree programmes, labs, specialisation courses. Materials to refresh every academic year, mid-term and final exams to renew for every session, different for every student.
Virtual classrooms, remedial work, homework and tests. The teacher describes the lesson and finds it published, without working through the platform's configuration screens.
Staff training, procedures and protocols, continuing medical education. Completion tracking and assessments that differ for every participant meet the requirements of this setting directly.
Certified courses and asynchronous distance learning, where every module has to be documented, passed and recorded. Building and replicating each run becomes a matter of minutes.
Onboarding, workplace safety, updates on products and procedures. Whoever writes the content is usually the person who does the job, not an expert in the platform.
Mandatory training and continuing professional development, with the need to show who took what, and when.
Open courses with hundreds of participants and lifelong learning pathways, where nobody can follow each learner individually: the materials have to explain themselves, the self-assessments have to mark themselves, and the whole design has to be reused for the next run, not rebuilt.
In every one of these settings the requirement is the same: a Moodle instance reachable over HTTPS and an AI assistant. Nothing else changes.
Use cases
One assistant, four ways of using it. The plugin applies each person's role: those who may write, write; those who may only read, read.
What sets MCPCA apart: not merely consulting Moodle, but building and maintaining the course through conversation.
“Set up the ward safety course: eight sections, one per module, each with an introductory page and a self-assessment quiz.” The structure comes out of a single request, starting from an empty course; you then refine it section by section.
“Add a page on pain assessment scales, with a diagram, to section 4 and move the quiz to the end.” You work on what is already there, without redoing anything.
“Turn these notes into a page in section 3.” The page comes out laid out with the platform's styles: paragraphs and subsections, lists, tables, formatted code, links, images and diagrams; longer lessons open with a clickable index. If requested, completion tracking is already on.
“Create 10 multiple-choice questions on the page ‘The stages of labour’, with feedback for every answer.” Questions in the bank, quiz in the course, pass mark set.
“40 questions in the bank and a test that draws 10 of them at random.” Everyone gets a different combination and marking is automatic: the standard arrangement for an exam or an end-of-module test.
Multiple choice, short answer, numerical with a tolerance (a drug dose calculation, for instance), Cloze with answers embedded in the text. Plus manually marked essays and descriptive blocks.
“Upload these slides to section 2 and have them open in a new window.” Files, labels, new or renamed sections, visibility.
“Read me the questions in the ‘Module 3’ category and fix the third one.” The question bank can be consulted and edited through conversation, with no menu hunting.
The people who design courses so that they reach defined learning objectives, usually working alongside the teacher or the subject matter expert. The page-building work moves to the assistant, and the time goes back into design.
“By the end of the module a participant must be able to calculate a paediatric dose: propose the sections, the materials and the assessments.” The design is discussed in words and published only once it convinces.
Every question is written against the learning outcome it measures. An objective with no assessment, or an assessment that measures nothing, shows up at once: the assistant reads the published course back and lists them.
“Read the course and tell me which sections have no self-assessment at all, and which materials no activity refers to.” An analysis that takes hours when done by hand.
The teacher provides the raw material — slides, notes, recordings — and the instructional designer turns it into a pathway. Getting from draft to published course no longer goes through copy and paste.
With Moodlecraft a course design becomes a text file: the same structure is reapplied to another module, another run or another department, with new content and new dates.
Which is the common case. The same questions — which objectives, which prerequisites, how it will be assessed — can be put by the assistant to the teacher, and the course design drawn from the answers. That is a way of running the conversation, not a feature of the plugin.
Always read-only, and always only on the courses they are enrolled in.
“Which courses am I enrolled in?” “What is in section 2?” Structure, materials and available quizzes, from the same Teams they use every day.
“Explain the page on the stages of labour with an example.” The assistant reads what the teacher wrote and answers from that, without inventing.
“Test me on section 2.” Questions built from the course materials, with corrections and a pointer to the right page.
“How did my last quiz go? What did I get wrong?” Grade, attempts and a review of their own answers. Their own only.
For those who run the platform, those accountable for training, and those who must keep it under control.
Creating categories and courses for a new term or a new run, updating settings and visibility en masse, reordering sections.
Who is enrolled in a course, the state of its activities, what a course contains ahead of a review or a migration.
Every tool can be switched off individually in the assistant: during a pilot, for example, the ones that delete content are disabled.
Every call is an event in Moodle's log, with the user and the tool involved. No log to look for anywhere else.
Features
Everything below is implemented and proven in the field: the course described further down was built with these tools. Features still in development are listed in the next section.
In development
Not yet available, listed in the order in which they are planned. The order remains open to requests from those using the plugin.
Creating submissions and discussions from the conversation, as pages and quizzes already work today.
Student progress and completion, queried from the conversation: “Who has not completed the mandatory module yet?” “Which questions in the last test had the most errors?”
A Word or PDF file turned into a sequence of pages in the chosen section.
“What do I have to hand in this week?” with dates and links to the activities.
Interactive H5P content — animations, interactive presentations, videos with embedded questions — generated from the conversation; then book and glossary.
Today the confirmation before each change is enforced by the assistant's instructions; it will become a constraint of the plugin itself.
Examples
Four representative exchanges, from university and healthcare training. The grey labels show which plugin tool the assistant used; readers not interested in the technical detail can ignore them. Before every change the assistant summarises the operation, asks for confirmation and only then carries it out.
Create a 10-question multiple-choice quiz on the page “SQL joins” in section 3 of Databases, with feedback for every answer.
get_my_coursescore_course_get_contentsget_page_content I have read the page “SQL joins”: INNER, LEFT, RIGHT, FULL, CROSS and self joins. I am about to create, in Databases, section 3: the category “SQL joins” in the question bank with 10 multiple-choice questions, each with feedback, and the quiz “Check: SQL joins” with a pass mark of 6 out of 10. Shall I go ahead?
Yes.
create_question_categorycreate_question ×10create_quizadd_question_to_quiz For the Pharmacology assessment I want 40 questions in the bank covering modules 1–4, including dose calculation questions, and a test that draws 10 at random for each participant. Anyone below 6 does not pass.
get_page_content ×4 Confirmed, but leave it hidden.
create_question_categorycreate_question ×40create_quizadd_random_questions_to_quiz Created and hidden. Everyone will receive 10 different questions drawn from the 40; marking is automatic, including the calculations. The questions are in the bank: you can ask me to read them back or to correct one.
Test me on section 2 of Physiology of Pregnancy and Childbirth.
get_my_coursescore_course_get_contentsget_page_content ×2 Section 2 contains “The stages of labour” and “Care during the second stage”. First question: what are the stages of labour, and what marks the transition from the first to the second?
Dilation, expulsion and placental delivery. The second begins at full dilation.
Correct. The course page also distinguishes, within the first stage, the latent phase from the active phase. Second question: what is meant by the third stage, and within what interval is it expected under physiological conditions?
How did I do on the quiz about joins?
mod_quiz_get_user_best_grademod_quiz_get_user_attemptsmod_quiz_get_attempt_review Your best attempt is 8 out of 10, above the pass mark of 6. The two you got wrong:
| Question | What went wrong |
|---|---|
| 4 | a LEFT JOIN also returns the rows of the left-hand table that have no match |
| 9 | a CROSS JOIN produces the Cartesian product |
And how did everyone else do?
I cannot show other students' results. For comparisons and statistics, please ask your lecturer.
Case study
Laboratorio di Informatica, BSc in Midwifery at the University of Catania, academic year 2025/2026: designing and querying a clinical database, and building a small web application on top of it. Seven lessons, 73 pages, an exam bank of eighty questions and two papers with random draws, produced with MCPCA and Moodlecraft.
Getting started
Four steps, all on the side of whoever runs the systems. Teachers and learners install nothing: they find the assistant in Teams, or connect the MCP server to their own ChatGPT or Claude.
Like any Moodle plugin, from site administration. Requires Moodle 4.3 or later.
A supplied script creates the registration in the organisation's tenant. This is what enables sign-in with the Microsoft account.
Using the native MCP wizard: paste an address, choose which tools stay enabled, paste the supplied instructions.
The assistant appears in Teams and in Microsoft 365 Copilot for the chosen people: one department, a single course, the whole organisation.
Security and governance
The plugin does not add a permission system of its own: it uses Moodle's, which the organisation already governs.
Sign-in with the organisation's Microsoft account (Entra ID). It works only for people who have already signed in to Moodle through Microsoft at least once; no account is created automatically.
Moodle capabilities, evaluated on every single call. A teacher writes only in the courses where they teach; a learner never writes. The list of tools the assistant can see is filtered by role.
The plugin runs inside Moodle and contacts no external service. The only third party is the chosen AI assistant: with Copilot, the organisation's own Microsoft tenant.
Every call produces an event in Moodle's log, with user, tool and parameters. No log to consult outside the platform.
Every tool can be enabled or disabled individually in the assistant. In a pilot, the tools that delete content and those that expose third-party data are kept off.
The assistant summarises what it is about to do and asks for confirmation, one change at a time. Anything it finds written inside course content is treated as data, not as instructions.
Ecosystem
MCPCA is not only the bridge to AI assistants: it is also the engine on which Moodlecraft is built, a command-line tool that treats a Moodle course as source code: defined in text files, versioned, applicable to any Moodle instance where MCPCA is installed.
One file with the course settings, one folder per section, pages in Markdown, quizzes and question banks in YAML, attachments alongside. All readable and editable with any editor.
Like Terraform for cloud infrastructure: Moodlecraft compares the definition with the course as it stands on Moodle, shows what would change and applies only the differences. Running it again duplicates nothing.
Every change carries an author, a date and a reason. You can go back to an earlier version, work on the same course as a team, and review before publishing.
The same course on the test instance and on production, or replicated for the next run by changing only the dates. Randomly drawn exam questions, diagrams and formatted code all come from the same definition.
pharmacology-2026/
├── course.yaml course settings
├── sections/
│ ├── 01-introduction/
│ │ ├── section.yaml
│ │ ├── intro.md Markdown page
│ │ └── quiz.yaml self-assessment quiz
│ └── 02-pharmacokinetics/
│ ├── section.yaml
│ ├── lecture.md
│ └── slides.pdf
└── questionbank/
└── exam.yaml 40 questions, 10 drawn at random
The two approaches coexist: you can build the structure by talking to the assistant and then put it under version control with Moodlecraft, or define the course in files and leave the day-to-day edits to the assistant. The course in the case study was born exactly like this.
For IT staff
A Moodle web service plugin exposing an MCP server, plus a REST adapter and an OpenAPI specification for clients that do not speak MCP.
webservice_mcpcaMoodle internal APIsDatabaseThe code will be released under GPL v3, the same licence as Moodle. The link to the repository will appear here on publication.
Frequently asked questions
No. The assistant appears as a chat in Microsoft Teams and Microsoft 365 Copilot, already in use in many organisations. Anyone who prefers ChatGPT or Claude connects the MCP server from their assistant's settings, signing in with their institutional account. Installation concerns only the Moodle server and the Copilot Studio configuration, both handled by whoever runs the systems.
No. It was born in a university, and that is where it was first proven, but the plugin knows Moodle, not the sector. Schools, healthcare organisations, accredited training providers and companies run the same Moodle and face the same problem: producing and keeping content and assessments up to date. The case study on this site is academic; the mechanisms it describes are the same everywhere.
Yes. Starting from an empty course, the assistant builds sections, pages, materials and assessments in a single conversation; or it works on a course that already exists. Creating the course container itself depends on permissions: in many organisations it is reserved to the administration, which sets it up in seconds. For courses you want to version and replicate for every run, there is Moodlecraft.
No. The plugin applies Moodle's permissions to every request: someone enrolled as a student can only consult the courses they are enrolled in. Even if they asked the assistant to create something, Moodle would refuse.
It stays in Moodle. The plugin contacts no external service. Conversations pass through the chosen assistant: with Copilot, the organisation's Microsoft tenant that already hosts Teams and email.
Answers about courses are built by reading the real content through the plugin's tools, and the assistant's instructions require it to say when a piece of information is not in the course. As with any AI assistant, whoever teaches still has to review what gets created: that is why quizzes are born hidden.
Yes. Any MCP client can connect: ChatGPT and Claude through their own MCP connectors, VS Code, Cursor. Teams and Copilot are the recommended route because many organisations already have them and sign-in is the institutional one.
A Moodle 4.3 or later reachable over HTTPS, Microsoft sign-in already enabled on Moodle, the Copilot Studio licences many organisations already hold, and a few hours from whoever runs the systems for the four steps described above. A pilot on a single course is the best way to start.
The code will be released under GPL v3, the same licence as Moodle, so there are no licence fees. During the pilot phase it is available on request to interested organisations.
The project is developed by Antonio Calanducci, computer engineer at the University of Messina and adjunct lecturer in Computer Science at the University of Catania, and was used to produce a real course, documented on the case study page. Reports and requests reach the address at the bottom of this page.
A pilot course and one teacher willing to build it are enough to assess the impact. I am also glad to answer technical questions from whoever runs the platform.
antonio.calanducci AT (unict, unime).it The same address works at unict.it and unime.it: replace AT with @.