schulcloud-mcp

Read-only access to a Schulcloud account — courses, boards, lessons, tasks and files — for Claude, via MCP, and for you, via a CLI that mirrors your coursework to disk.

Both are front ends over one core library and one live Schulcloud session, kept alive on a Pi.

Built and verified against schulcloud-thueringen.de with a live student account. Everything in docs/API.md was confirmed against the running instance, not inferred from the upstream source.

What Claude can do with it

"What do I have due this week?" "Find the material about Verschlüsselung and explain the Caesar cipher worksheet." "Summarise the routing lesson from the LF10 course."

Twenty-four tools, all read-only:

whoami	account, school, roles — also a connectivity check
list_courses	all courses, with ids
get_dashboard	the tiles as pinned on the web dashboard
get_course	one course's boards, topics and tasks
get_board	a column board in full: columns, cards, text, links, files
get_lesson	a topic's text sections, materials, files and tasks
list_tasks	homework across all courses, by due date
get_task	one task: description, due date, attachments, and your submission
list_submissions	what you handed in, and what is still ungraded
list_files	files attached to a board element, topic, task or submission
download_file	fetch such an attachment and extract its text, or view an image
fs_list	list a folder of the file manager — Persönliche, Kurs-, Team-, Geteilte Dateien
fs_tree	everything below a file-manager folder, as a tree
fs_find	find file-manager files and folders by name
fs_read	read a file-manager file, extracted like download_file
search	keyword search across everything — including the text inside PDFs and Office files
refresh_index	re-read Schulcloud now, per course or in full
what_changed	what appeared, changed or vanished since a date
index_status	how fresh the index is
list_rooms	rooms ("Räume"), which are a separate space from courses
get_room	one room: its boards, members and what you may do there
list_classes	classes ("Klassen") with their teachers, and group membership
list_news	school and course announcements
api_get	GET-only escape hatch for uncovered API surface

download_file extracts text from PDF, DOCX, XLSX, PPTX and OpenDocument files and returns images inline for Claude to look at. Image-only PDFs — scans with no text layer, which are common in this account — are reported as such rather than as an empty result.

Attach a course instead of asking for it. Every course and room is also an MCP resource (schulcloud://courses/<id>, schulcloud://rooms/<id>) holding the same overview get_course and get_room return. In Claude Code, type @ and part of the course name.

Two ready-made prompts, in German because the school is:

zusammenfassung kurs [fokus]	summarise a course or room: topics, tasks, key material, each with its source
pruefungsvorbereitung kurs [thema] [datum]	prepare for an exam: scope, explanations, practice questions, a study plan

In Claude Code they run as /mcp__schulcloud__zusammenfassung Mathe_10b. Claude Code splits arguments on spaces and drops extra words, so join words with _ (Lineare_Funktionen) and skip an optional argument with -. kurs is any unambiguous part of a course or room name, or its id.

The CLI

schulcloud login --server https://mcp.example.org --token <token>
schulcloud sync --dry-run      # see what would be mirrored
schulcloud sync                # mirror coursework to ~/Schulcloud
schulcloud refresh --course <id>
schulcloud fs tree /courses    # browse the file manager ("Dateien")
schulcloud fs get "/courses/<course>/<folder>"
schulcloud token set           # the monthly chore: hand the Pi a fresh Schulcloud token

It talks only to the Pi and holds no Schulcloud credential — see docs/CLI.md. A fresh token can also be pasted into the server's /token page; either way the server checks it with Schulcloud and swaps it in without a restart.

Quick start

cp .env.example .env        # fill in TSC_URL and TSC_JWT_COOKIE
npm install
npm run build
npm run probe               # verifies the token and API against the live instance

To try it on your own machine — Docker stack, Claude Code, and the CLI — follow docs/LOCAL.md. To put it on a Pi behind Caddy and a VPS, and connect claude.ai, follow docs/PI.md; docs/DEPLOYMENT.md explains the pieces.

Getting TSC_JWT_COOKIE takes four clicks in DevTools and then lasts 30 days — provided you close the Schulportal window afterwards. See docs/AUTH.md; that caveat is not optional.

Design decisions

Bearer token plus a keepalive. The instance's jwt cookie works verbatim as Authorization: Bearer — no cookie jar, no connect.sid. Its exp claim (30 days) is only a ceiling: the real limit is a 2-hour server-side session TTL that any request slides, so the server calls refresh-session every 30 minutes (the one non-GET request here, and not exposed as a tool).

The sharp edge is subtler and cost two endurance tests to find: the cookie you copy is the browser's own session token, so a Schulportal tab left open will auto-logout after ~2 hours and revoke this server's token with it. Copy the token in a private window and close it. See docs/AUTH.md.

claude.ai reaches it by a secret path, for now. Its connector dialog takes a URL and no header, so MCP_PATH_SECRET serves MCP at /<secret>/mcp without the bearer token — never logged, redacted by the Caddy snippet, and a stopgap until the endpoint speaks OAuth. Claude Code and the CLI keep the bearer token. See docs/DEPLOYMENT.md.

Read-only by construction. Every method on the API client is a GET, including api_get. The endpoint is internet-facing by necessity (Claude's connectors call it from Anthropic's cloud), so the fact that a leaked token cannot be used to act as the user is the main safety property. Adding one write tool would forfeit it.

An index, with an honest bypass. Postgres holds crawl generations and a german + pg_trgm full-text index over extracted file text, so search covers the inside of PDFs rather than just their names. Every result states how fresh the index is, and fresh=true bypasses it for a live read — an agent should never be quietly misled by stale data. Without DATABASE_URL the server still works: search falls back to crawling live.

Generations, not timestamps. Sync cursors and change detection compare crawl generations by identity. Measured: the course-board endpoint returns request time as updatedAt, so a timestamp cursor would report everything as changed on every crawl — and could never detect deletions.

Assembled, not raw. get_board makes three kinds of upstream call and stitches the results — board skeleton, card bodies, and a files-storage lookup per file element — because a model asking "what's on this board" wants the answer, not a traversal plan. Output is Markdown with ids preserved for follow-up calls, not raw JSON.

Possible extensions — full-text search over file contents, a cache with a bypass, "what's new since…" — are sketched with their trade-offs in docs/ROADMAP.md. None are built.

Layout

src/
  core/         client, types, board assembly, crawler, extraction, paths
  store/        Postgres: crawl generations, diffs, full-text search
  indexer/      crawl → persist → mirror bytes → extract text → index
  mcp/          MCP server, tools, resources and prompts
  http/         express app, bearer auth, /api for the CLI
  cli/          CLI config, API client, sync engine
  bin/          http, stdio and cli entry points
docs/           API findings, auth, deployment, CLI, roadmap
deploy/         Caddyfile snippet, the Pi's compose file
scripts/        probe, smoke, session diagnostics
vendor/         upstream clones, git-ignored, for reference only

core/ knows nothing about MCP, HTTP or the CLI: it holds the Schulcloud client, the traversal every feature needs, document extraction, and the path sanitisation that both the server's mirror and the CLI's sync depend on.

Development

npm run dev        # watch mode, runs src/ directly
npm test           # unit tests, no network
npm run probe      # check assumptions against the live instance
npm run smoke      # full end-to-end: real server, real client, real data
npm run session-diagnose  # instrument what actually ends the session (~2.5h)
npm run keepalive-status  # is the deployed container holding its session?
npm run publish-image     # build amd64 + arm64 and push to registry.mc02.dev — the Pi only pulls
npm run typecheck

npm run smoke starts the HTTP server, connects a real MCP client over Streamable HTTP and exercises every tool against the live account — 74 checks (76 with the index) covering the auth gate and the secret path, the protocol handshake, every content chain, file extraction, resources and prompts, token replacement, api_get's guard rails and error handling.

Upstream

Reference clones live in vendor/ (git-ignored):

git clone --depth 1 --filter=blob:none https://github.com/hpi-schul-cloud/schulcloud-server.git vendor/schulcloud-server
git clone --depth 1 --filter=blob:none https://github.com/hpi-schul-cloud/file-storage.git      vendor/file-storage
git clone --depth 1 --filter=blob:none https://github.com/hpi-schul-cloud/nuxt-client.git       vendor/nuxt-client

Most of that organisation's ~100 repositories are archived or superseded; those three are the live ones that matter. The instance's own OpenAPI documents (/api/v3/docs-json, /api/v3/file/docs-json) are more authoritative than any of them — see docs/API.md.