Brightspace-MCP

An MCP server that can expose your Brightspace account as tools available for an LLM to call.

Some assignments confound me in format and feed because every professor has a different way of assigning things, being a quiz, a content item, something on the course calendar, or even something else. Each feed is individually incomplete when calling the API so there's availability to call batch functions for multi entry retrieval.

Tools

This section was written by Claude from the tool docstrings in server.py.

All course-scoped tools take an orgid (the OrgUnit ID) from getClasses. getBatch* variants take a list of orgids — a real array, a JSON-array string, or a comma-separated string, since some MCP clients stringify array args — and fan out across courses concurrently.

Account

• getUser — the calling user's profile (whoami).

• getClasses — current enrollments. Pass false for a de-duped, current-term-only list (id / name / code); pass true for the raw D2L enrollments feed including past and hidden courses. Prefer false to save tokens.

Grades

• getAssignedGrades — the user's grade values for one course.

• getAllGrades — the full grade objects (definitions + values) for one course.

What's due

• getWeeklyTodo / getBatchWeeklyTodo — content-linked items due in the next days (default 7). Each item has ItemName, DueDate, ItemUrl.

• getAllDueItems / getBatchAllDueItems — content-linked items with no date window (past-due through everything upcoming).

• getWeeklyCalendarEvents / getBatchCalendarEvents — course calendar events in the next days (default 7), each with a CalendarEventViewUrl.

• getQuizzes / getBatchQuizzes — every quiz in the Quizzes tool, whether or not an instructor linked it into content. Inactive quizzes are hidden unless include_inactive=true.

• getEverythingDue / getBatchEverythingDue — the merged content + quizzes + calendar picture, deduped by GradeItemId (then normalized title). Each entry carries a Sources list naming the feeds it came from. days (default 180) bounds only the calendar window. Use this when nothing can be missed.

Content

• getCourseToc — the course table of contents. Default is a flat list of topics (module path, Type, Url, TopicId, GradeItemId, availability fields); full=true returns the raw nested D2L tree. The TOC has no due dates or completion state.

• getTopicFile — downloads the file backing a File-type content topic (GET /le/1.82/{orgid}/content/topics/{topicId}/file). Takes orgid plus a topicId from getCourseToc (only works where the topic's Type is a file, not a link/URL topic). Returns ContentType / FileName / Size, then Text for HTML/text/XML/JSON/CSV payloads or Base64 for binary ones (PDFs, slide decks, images). Redirects to D2L's signed storage URLs are followed.

• getSyllabus — the course syllabus / content overview (GET /le/1.82/{orgid}/overview). Returns Description (rich text) and HasAttachment; when an attachment exists it is downloaded too and returned under Attachment in the same ContentType / FileName / Size / Text or Base64 shape as getTopicFile. Pass include_attachment=false for just the metadata. Not every course uses the overview — if it comes back empty, look for a "Syllabus" File topic via getCourseToc + getTopicFile.

Completed content items and inactive quizzes are filtered out by default; include_completed / include_inactive bring them back. Calendar events carry no completion state, so a calendar-only entry can't be filtered that way — cross-check getAssignedGrades for submission/score status.

On any upstream failure a tool returns {"error": ..., "endpoint": ...} in place of its normal payload (rather than a bare null); a 401/403 — usually expired session cookies — adds a "hint" that says so.

getLink (Kaltura lecture-video transcription) is registered only when the optional transcription extra is installed; getLTILink (LTI quicklink redirects) is checked in but commented out. See Optional tools.

Related MCP server: Canvas Assignment Assistant

Architecture

Claude / MCP client
      │  HTTPS  + Authorization: Bearer <token>
      ▼
nginx  ($MCP_PUBLIC_HOST, TLS termination) ← My setup, http streamable claude requires HTTPS so I used Cloudflare
      │  HTTP, Host preserved
      ▼
brightspacemcp  (streamable-http, 127.0.0.1:8008)   ← this repo
      │  session cookies + browser UA
      ▼
purdue.brightspace.com/d2l/api

• Transport: streamable-http bound to loopback. TransportSecuritySettings pins allowed_hosts / allowed_origins (from MCP_PUBLIC_HOST, default mcp.xennick.com) so the SDK's DNS-rebinding protection accepts the Host nginx forwards.

• Inbound auth (MCP client → this server): the RequireToken middleware rejects any request without Authorization: Bearer <MCP_INBOUND_TOKEN> (constant-time compare); the token lives in .env and is minted by mint_token.py.

• Outbound auth (this server → Brightspace): auth.return_cookies() reads d2lSessionVal / d2lSecureSessionVal from .env (via python-dotenv). Those are a logged-in browser session's cookies; a separate process is expected to refresh them into .env. Requests also send a desktop-browser User-Agent.

• Deploy: deploy/brightspace-mcp.service runs the brightspacemcp console script under systemd with uv run --frozen --no-sync (never touches the lockfile at boot) and a strict sandbox (ProtectSystem=strict, ProtectHome=read-only, restricted address families, private tmp).

Caveat

My inbound side has self-generated token auth at the moment and the outbound side talks to the D2L api with a scraped session cookie, not an OAuth 2.0 app.
"Unfortunately, we have not yet provided students with OAuth 2.0 access to Brightspace for personal use."
Because of this, it is single-user, tied to a single account, breaks on session expiry, and is likely against API terms. It is a personal tool, not a multi-user service. A future production version would also be able to register a D2L app and use the OAuth flow for the Brightspace call.

Optional tools

getLink — launches the course's Kaltura LTI in a headless Playwright browser, grabs the video's index.m3u8, and runs it through openai-whisper for a transcript. Its dependencies live in the transcription extra (playwright + openai-whisper, which pulls PyTorch), kept out of the base install because they're large and the transcription itself is slow on a small host.

• The base server imports fine without them; server.py checks importlib.util.find_spec at startup (_HAS_TRANSCRIPTION) and only calls mcp.tool()(getLink) when both are present, so a lean install advertises exactly the tools it can run.

• Called without the extra (e.g. registered by hand), getLink returns {"error": "transcription extra not installed", "hint": ...}.

Enable it with:

uv sync --extra transcription
playwright install chromium

getLTILink — resolves /d2l/common/dialogs/quickLink/... redirects (a thin wrapper over a raw authenticated GET). Checked in with its @mcp.tool() line commented out; un-comment to register it (no extra dependencies).

Setup

Requires uv. Python ≥ 3.14 is fetched by uv automatically — you don't need it installed already.

Guided (recommended)

./setup.sh

It's interactive and safe to re-run. It will:

1. install uv if it's missing (offers to run the official installer),

2. uv sync (and, if you say yes, the heavy transcription extra),

3. prompt for your two Brightspace session cookies,

4. mint the inbound bearer token via mint_token.py,

5. write .env, and

6. print the line to paste into your MCP client:

   Copy this whole token for input: "Bearer <token>"

Then run the server:

uv run brightspacemcp          # or: python -m brightspacemcp

It listens on http://127.0.0.1:8008. Point an MCP client at that (directly, or through a TLS proxy as above).

Manual

uv sync

The base uv sync is lean — mcp, httpx2, python-dotenv. See Optional tools for the transcription extra.

Create .env with a current browser session's cookies:

d2lSessionVal=...
d2lSecureSessionVal=...
MCP_INBOUND_TOKEN=...        # bearer token MCP clients must send
MCP_PUBLIC_HOST=...          # optional; public hostname nginx serves (default mcp.xennick.com)

Copy the d2l* cookies from your browser's dev tools while logged into purdue.brightspace.com (DevTools → Application / Storage → Cookies — the entries named d2lSessionVal and d2lSecureSessionVal). For the token, run

uv run python mint_token.py

which generates one, upserts MCP_INBOUND_TOKEN into .env (leaving your other lines alone), and prints the Bearer … string. --show reprints the current one; --force replaces it.

Then run the server as above.

Refreshing expired cookies

When tools start returning 401/403, your d2l* cookies have expired. Re-run ./setup.sh (press enter to keep the token and re-paste the two cookies), or edit the two d2l* lines in .env by hand.

Deploy as a service

cp deploy/brightspace-mcp.service /etc/systemd/system/
systemctl daemon-reload
systemctl enable --now brightspace-mcp.service

Adjust User, WorkingDirectory, and the uv path in the unit first. After changing dependencies, run uv sync by hand — the unit deliberately runs --frozen --no-sync.

Layout

src/brightspacemcp/
  __init__.py   exports main()
  __main__.py   python -m brightspacemcp
  server.py     MCPServer, all @mcp.tool() definitions, request helpers
                (JSON / file download), feed-merge logic, RequireToken
  auth.py       Brightspace session cookies from .env (outbound)
setup.sh        interactive first-run setup (deps + .env + token)
mint_token.py   generate MCP_INBOUND_TOKEN and upsert it into .env
deploy/
  brightspace-mcp.service
pyproject.toml  uv_build, src layout, `brightspacemcp` console script