ADAM MCP
Uses a local Chrome session for authenticating to ADAM and fetching course pages, with an option to run headless after an initial login.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
ADAM MCP
ADAM MCP connects MCP-compatible AI clients to your University of Basel ADAM courses.
It runs on your machine and uses a dedicated Chrome session to access the ADAM courses available to your account. An MCP client can list courses, find materials, read approved text, and surface deadlines. See client compatibility for configuration support and verification status.
This is a community project, not a University of Basel service.
0.4.0 is in preparation, not published. It adds selected Web Link metadata, session ownership and boundary hardening. 0.3.0 is published, including selected local downloads and discovery efficiency improvements. See the release and public-package verification record.

The narrated overview illustrates a study workflow with realistic synthetic course data and scripted responses. It does not record a supported client connected to live ADAM or establish client compatibility.
What you can ask
Show timestamped course news from this week.
What deadlines do I have?
Find the exercise sheet about Fourier transforms.
Summarize this course page and give me the ADAM link.
List the material available for this course.
Answers include canonical ADAM URLs (https://adam.unibas.ch/go/...).
Related MCP server: Quirkus
Quick start
You need Node.js 22+ and Google Chrome. Use a supported Node.js LTS release.
Install the published package. Guided setup checks your installation, configures one client with a backup, opens Chrome for sign-in, and verifies your course listing.
npm install -g adam-mcp@0.3.0
adam-mcp setup --client codex --write --loginReplace codex with cursor, claude-desktop, windsurf, or vscode. For VS Code, supply --output with the intended workspace's .vscode/mcp.json path. Restart your client, enable adam, then ask: What courses am I in? Include their ADAM links.
Preview without changing settings: adam-mcp setup --client codex. Diagnose a connection: adam-mcp doctor --client codex. Credentials stay in the Chrome sign-in window. Setup never asks for them in the terminal or chat.
See client compatibility for what has been verified and the setup guide for manual configuration, backups, and recovery.
Run without a global install
The published package ships a self-contained CLI:
npx -y adam-mcp@0.3.0The exact npm version was installed and checked outside the repository. See the publication record for source, integrity, client evidence and limits.
Build from source
git clone https://github.com/Arudchayan/adam-mcp.git
cd adam-mcp
npm ci
npm run setup -- --client codex --write --loginThe repository commands below also work from a source checkout. Installed users can use adam-mcp login, adam-mcp status, and adam-mcp logout directly.
Session management
Without a client argument, interactive setup asks which client to use; noninteractive setup prints all supported snippets. Setup previews by default. Add --write to install the selected configuration with a backup, and --login to sign in and verify courses.
To sign in separately, run
npm run loginand complete ADAM sign-in in Chrome. The window closes after verification; a background Chrome process keeps the session available. After login, checkadam_session_status, then calladam_list_courses.
Restart the client, then ask:
What courses am I in? Include the ADAM URL for each.
Login reuses a valid session without opening Chrome. Close the sign-in window or
press Ctrl+C to cancel; failed attempts close their window. Use
npm run login -- --force for a fresh window, or npm run login -- --timeout-ms 120000 for a
two-minute sign-in wait. npm run logout stops the local session and disables
automatic reconnection until explicit login. npm run logout -- --purge also
removes the dedicated profile; local logout does not sign you out of SWITCH in
other browsers.
Host-specific files (Cursor, Claude Desktop, VS Code, Windsurf, Claude Code, ChatGPT desktop, Codex): docs/setup.md.
Save a selected file locally (0.3.0)
Find the file ID using adam_list_files, then run this command in your terminal.
Use an existing directory outside any Git checkout:
adam-mcp download 100011 --output-dir "D:/Study/ADAM" --confirm100011 is an example; replace it with your selected file ID. This explicitly
saves that file to your computer without sharing its bytes with the model.
The CLI returns its path, source URL, size and SHA-256. Duplicate names receive
a suffix; existing files are preserved. Ctrl+C cancels the transfer. Downloads
are limited to 256 MiB and 120 seconds; empty and HTML-shaped responses are refused.
The destination filesystem must support hard links. This adds no MCP download
tool and does not approve text extraction. Details and cleanup.
How it works
The client starts ADAM MCP as a local process and communicates over stdio. ADAM MCP reads through your dedicated Chrome session at https://adam.unibas.ch. Authentication stays in Chrome; no hosted MCP endpoint is provided.
Architecture · What is in scope · Capabilities · Roadmap
Tools
Tool | Returns |
| Enrolled courses and ADAM URLs |
| Course snapshot and folder contents |
| Page text ( |
| File metadata and URLs, not bytes |
| Local PDF, text, notebook ( |
| Title matches in enrolled objects |
| Exercise deadlines, Dashboard tasks and events, dated course news |
| One entry per assignment with deadline and your status; |
| Forum meta and thread summaries; post bodies need |
| Chrome session in the live runtime |
Document and thread text requires approval for the selected object. Your connected model receives the returned text. See tool contracts for confirmation, pagination, recovery and completeness rules.
For page, file-text, exercise, and forum-post reads, request the specific object first, then approve the read. Clients with elicitation support can ask for approval interactively; other clients pass confirm: true. Declining or cancelling returns no body text. Forum summaries do not require body-read approval. See tool contracts for the precise contract.
Limitations
No writes, mail, grades, or exam (
tst) objects.Search covers titles and available page text in your enrolled course tree. It does not use ADAM’s global search, and search does not open exercise pages. Calendar dates come from exercise deadlines, the Dashboard Tasks and Calendar blocks, and dated course text. File upload times are not reported as events, and the full ILIAS calendar interface is not read.
Scanned PDFs are not OCR’d.
Supported PDFs are parsed page by page, with an 8 MiB transfer cap, 40,000 text characters, and at most 20 pages per call (default eight). Pass
startPage(or the returnednextStartPage) to read later pages. Empty/malformed files fail closed. Formula interpretation and exact visual layout are not guaranteed.Course pages and complete cross-course walks refresh after one minute. Restart the MCP server for an immediate refresh. Existing pagination retains its original snapshot for ten minutes; expired cursors require restarting the list.
Cross-course reads have a two-minute elapsed budget and report partial results when reached. The browser queue admits at most 32 waiting operations and limits waiting to 30 seconds. Cancelling a request releases its work when nobody else needs it; cached reads reuse navigation and parsing.
Version 0.3.0 also shares complete discovery walks within each enrolled course. PDF parsing admits at most eight waiting requests and 32 MiB of waiting input; queued requests expire after 30 seconds. Overload returns a retryable error.
Search, calendar and news accept an enrolled
courseRefIdto narrow discovery. Userefresh: truewithout a cursor for fresh results.snapshotAtidentifies the listing snapshot; each item's provenance records when it was fetched. Partial results include coverage reasons.News with a cutoff includes only items with a known timestamp. Undated news remains visible without a cutoff. This does not compare file inventories; no matching news is not proof that nothing changed.
Deadline clock times without an offset use Europe/Zurich. Ambiguous or invalid dates/times are omitted. Date-only evidence uses UTC midnight as a representation, and does not establish a midnight deadline.
Privacy
Tool results go to the model your client is already using. Course content is copyrighted.
Develop
npm test
npm run typecheck
npm run check # tests, typecheck, build, built-package verification, dependency audit
npm run release:check # also verifies a clean tarball installation
npm run benchmark:reads # synthetic cached-read timing; no live ADAM access
npm run benchmark:discovery # cold/scoped/large trees and queue pressure; syntheticCONTRIBUTING.md · Release guide · AGENTS.md
License: GPL-3.0-or-later.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Absorb LMS MCP server for ChatGPT, Claude and any AI agent: list users, list course enrollments, list overdue learners, approve enrollment request, enroll users in a course, message learners, create user and deactivate user in your own Absorb LMS account. Cotlet works in the account through your Cloud Browser, a real Chrome where you sign in to Absorb LMS yourself; Cotlet does not store your password and you need no Absorb LMS API key. About the official Absorb LMS API: Absorb LMS has an Integration API, a REST API with JSON for learners, courses, departments and grades, but it needs a private key tied to your portal that your Absorb account team provides, and your own code (checked 20 Oct 2026). Absorb's Aura MCP for Claude and ChatGPT is planned for 27 Oct 2026 and is read and answer only. Setup and FAQ: https://cotlet.dev/integrations/absorb-lms-api-mcp-server
Zotero MCP server for ChatGPT, Claude and any AI agent: add item by DOI, search library, add to collection, get annotations, create note and export citations in your own Zotero account. Cotlet works in the account through your Cloud Browser, a real Chrome where you sign in to Zotero yourself; Cotlet does not store your password and you need no Zotero API key. About the official Zotero API: Zotero has a public Web API v3 with read and write access (checked 16 Oct 2026). You create a key in your zotero.org account settings and send it with each request. Setup and FAQ: https://cotlet.dev/integrations/zotero-api-mcp-server
Web search, data extraction and controlled browser automation for AI assistants via OAuth.
Your personal data for AI — Telegram, bank, courses, Zoom & more, scoped to you.
Related MCP Servers
- AlicenseAqualityBmaintenanceEnables AI assistants to read a student's own Blackboard (Ultra) data—courses, assignments, grades, announcements, and attachments—through their existing logged-in browser session, read-only and locally.109MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI assistants to read courses, upcoming deadlines, assignment details, grades, announcements, modules, files, inbox, and timetable from U of T Quercus and ACORN through read-only tools.1MIT
- AlicenseNot gradedqualityCmaintenanceLets AI assistants read a student's Canvas LMS/Moodle courses, deadlines, announcements, grades and course files using their normal browser login instead of an access token. It is read-only and keeps all session data and downloads on the user's own machine.36 npm1MIT
- AlicenseBqualityCmaintenanceConnects AI assistants to the UNED virtual campus (agora.uned.es, Moodle) so students can log in via SSO, sync course materials, and search or read extracted text with page-level citations. It also surfaces deadlines, announcements, grades, classes and calendar events, tracks topic-by-topic progress, generates weekly study plans and exports dates to an .ics calendar.211MIT