Skip to main content
Glama

waterloo-learn-mcp

MCP server exposing your Waterloo LEARN courses to AI apps. LEARN sits behind WatIAM + Duo, so auth is split out: you log in once in a real browser, the session is saved to auth.json, and the server reuses those cookies over plain HTTP until they expire. (A headless browser is only launched for rare scrape fallbacks, so npx playwright install chromium is still part of setup.)

auth.json, oauth.json, and .env.local hold secrets. They are gitignored — treat them like passwords.

Install

npm install
npx playwright install chromium
npm run build
npm run login          # browser opens; sign in + approve Duo. Saves auth.json

Optional: npm run login can autofill your WatIAM username/password before waiting for Duo. Add both values to .env.local or export them in your shell:

WATIAM_USERNAME=your-watiam-user
WATIAM_PASSWORD=your-watiam-password

Related MCP server: Horizon

Tools

Tool

Args

Returns

list_courses

—

Course names + ou IDs

get_announcements

courseId

Announcements (title, body, date, attachments)

get_content

courseId, query?

Content modules/topics with URLs; query returns only matching hierarchy paths

get_topic_file

courseId, topicId, pages?

Adaptive whole-deck contact sheets, or detailed slide images when pages is supplied

get_grades

courseId

Grade items (grade, points, weight, feedback)

get_assignments

courseId

Assignments with due dates, instructions, your submission status + files, released feedback

get_upcoming

courseId, daysAhead?

Due dates / events (default 30 days)

get_course_outline

courseId

Official course outline/syllabus text from the local cache, refreshing from Outline.uwaterloo.ca when missing

Connect to Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json, then restart the app:

{
  "mcpServers": {
    "waterloo-learn": {
      "command": "node",
      "args": ["/absolute/path/to/waterloo-learn-mcp/dist/index.js"]
    }
  }
}

Local, no tunnel, no exposed session. Details: skills/connect-claude-desktop/SKILL.md.

Connect to a web chat with ngrok

ChatGPT and Claude.ai are cloud-hosted, so they need a public HTTPS URL for this MCP server. The default path is ngrok: the MCP server runs on your laptop, and ngrok publishes it at a stable HTTPS dev domain.

First, install and sign in to ngrok:

brew install ngrok/ngrok/ngrok
ngrok config add-authtoken <your-ngrok-authtoken>

Get your free dev domain from the ngrok dashboard, then run the one-time setup:

npm run setup:ngrok

When prompted, paste the domain only, for example example.ngrok-free.app. The script saves it to .env.local, starts the local HTTP server, starts ngrok, and prints your connector URL:

https://<your-ngrok-domain>/mcp

Use OAuth. Leave OAuth Client ID and Client Secret blank. When the authorization page opens, paste the connection code printed by the setup script (LEARN_MCP_TOKEN from .env.local).

  • ChatGPT → Settings → Connectors → Developer mode → Add custom connector → Server URL.

  • Claude.ai → Settings → Connectors → "+" → name + URL.

OAuth clients and tokens are saved in oauth.json so connectors keep working after server restarts. Treat LEARN_MCP_TOKEN like a password. The legacy URL https://<your-ngrok-domain>/mcp/<secret> still works for clients that cannot use OAuth.

After a reboot with ngrok

Your ngrok domain, MCP token, OAuth clients, and LEARN session all persist. After restarting your computer, run:

npm run start:ngrok

Leave that terminal running while you want ChatGPT or Claude.ai to reach LEARN.

If the LEARN session expired, start:ngrok exits before starting the tunnel. You can also check it directly:

npm run check:auth

To check the tunnel from another terminal:

curl -s -H 'ngrok-skip-browser-warning: true' https://<your-ngrok-domain>/health
curl -s -H 'ngrok-skip-browser-warning: true' https://<your-ngrok-domain>/.well-known/oauth-authorization-server

Expected:

  • /health -> ok

  • OAuth metadata -> JSON with a registration_endpoint

ngrok may show a browser warning during OAuth. Click through once. Command-line checks can skip it with the ngrok-skip-browser-warning header shown above.

Tailscale alternative

Tailscale Funnel also works if you prefer it. It has a nice reboot story because Funnel persists its tunnel config; only the local HTTP server needs to be running.

npm run setup:tailscale
npm run start:http
npm run autostart:http  # optional: start HTTP server on login
npm run stop:http       # stop HTTP autostart and free port 8787

Use this connector URL:

https://<device>.ts.net/mcp

After a reboot with Tailscale:

  • Ran autostart:http -> nothing to do; it restarts itself on login.

  • Didn't -> npm run start:http to bring the server back.

If Claude/ChatGPT cannot fetch OAuth config or cannot connect to <device>.ts.net:443, reset Tailscale Serve/Funnel:

tailscale funnel reset
tailscale serve reset
tailscale funnel --bg 8787

Notes

  • list_courses uses the enrollments API, falling back to homepage scraping. Official titles from Outline.uwaterloo.ca are cached in cache/outlines/viewer-outlines.json and reused when the outline session is expired. Other tools call D2L's REST API through the authenticated session.

  • get_course_outline reads cache/outlines/ first. Cached outlines are checked against the published revision date and automatically refetched when the instructor publishes a new revision. If a course is not cached, it checks Outline.uwaterloo.ca's enrolled-course viewer, then falls back to outline links posted in LEARN content. If neither exists, look for an uploaded outline/syllabus PDF in get_content.

  • get_content(courseId, query="Week 13 Meeting 1") searches module paths and topic titles before returning data, avoiding a full-semester tool response when only one item is needed. Omit query for the complete hierarchy.

  • get_topic_file returns native MCP images so the model can read diagrams, not just text. Without pages, decks of up to 4 slides return individual PNGs; larger decks return numbered JPEG contact sheets with up to 12 slides each (75-slide overview cap). Pass pages="4", pages="2-6", or pages="2,4,7-9" for detailed 800px PNGs (30-slide cap per call).

  • Downloaded PDFs and converted PowerPoints are cached in-process for 15 minutes, up to 100 MB total, so overview-to-detail follow-ups avoid another LEARN download or LibreOffice conversion. PDFs need nothing extra; PowerPoint topics need LibreOffice (brew install --cask libreoffice). Works in Claude (Desktop + Claude.ai) and ChatGPT.

  • "No valid LEARN session" (or tools failing after weeks) = session expired → npm run login again. Independent of reboots.

  • Set LEARN_SKIP_OUTLINE_REFRESH=1 in .env.local to keep startup-triggered logins from opening Outline.uwaterloo.ca or rebuilding the outline cache. Existing cached outlines are still used; live outline lookups can still refresh individual entries when requested.

  • Override with env vars: LEARN_BASE_URL, LEARN_AUTH_FILE, LEARN_OUTLINE_CACHE_DIR, LEARN_SKIP_OUTLINE_REFRESH, LEARN_TOPIC_FILE_CACHE_TTL_MS, LEARN_TOPIC_FILE_CACHE_MAX_BYTES, PORT, LEARN_MCP_TOKEN, WATIAM_USERNAME, WATIAM_PASSWORD, WATIAM_LOGIN_DOMAIN.


Disclaimer

This is an unofficial, independent tool built for personal academic use. It is not affiliated with, endorsed by, or supported by the University of Waterloo, D2L, or any related entity. "Waterloo LEARN" and related marks belong to their respective owners.

The server acts on your behalf using your own credentials to access only your own course data — it accesses nothing you couldn't already see by logging into LEARN yourself. You are responsible for using it in accordance with the University of Waterloo's acceptable-use policies and LEARN's terms of service. Don't use it to access data that isn't yours, and don't share course content in ways that violate copyright or your instructors' wishes.

LEARN's internal APIs and page structure can change without notice, which may break this tool at any time. It is provided as-is, without warranty of any kind. AI models can also misread or hallucinate content — always verify grades, due dates, and other important information against LEARN directly. Use at your own risk.

Available Tools

5 tools
get_announcementsGet AnnouncementsA

Get announcements/news posted by instructors for a course. Returns title, body, posted date, and attachments. Use to answer: "Any new announcements?", "What did the professor post?"

Inputs: courseId (required).

get_contentGet Course ContentA

List the content modules and materials (lectures, slides, files, links) for a course as a nested table of contents. Use to answer: "What materials are in week 3?", "Where are the lecture slides?"

Inputs: courseId (required).

get_gradesGet GradesA

Get your grades for a course: each grade item with displayed grade, points, weight, and feedback. Use to answer: "What are my grades?", "How did I do on the midterm?"

Inputs: courseId (required).

get_upcomingGet Upcoming Due DatesA

Get upcoming calendar events and due dates (assignments, quizzes, exams) for a course. Use to answer: "What is due this week?", "When is the next deadline?"

Inputs: courseId (required), daysAhead.

list_coursesList CoursesA

List the courses you are enrolled in on Waterloo LEARN. Returns course names and their ou (org unit) IDs, which the other tools take as courseId. Use to answer: "What courses am I taking?"

Inputs: no named parameters.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 5 tools

    Disambiguation5/5

    Each tool targets a distinct aspect of the LEARN experience (announcements, content, grades, upcoming events, courses), with no overlaps in purpose.

    Naming Consistency5/5

    All tool names follow the consistent pattern get_<resource> in snake_case, making them predictable and easy to remember.

    Tool Count5/5

    5 tools is an ideal size for a student-focused LMS server, covering all essential read operations without unnecessary bloat.

    Completeness5/5

    The toolset provides comprehensive coverage of a student's common needs: viewing courses, grades, materials, announcements, and upcoming deadlines. No obvious gaps.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers