EssenceTeacher MCP

An MCP server that puts a teacher's own courses, assignments, submissions and student-support queue in front of an AI assistant — so "where is my class stuck?" is answered from the real data instead of from a description of it.

Companion to essencescholar-mcp, which does the same for the research platform.

Install

Nothing to install — run it with uvx:

{
  "mcpServers": {
    "essenceteacher": {
      "command": "uvx",
      "args": ["essenceteacher-mcp"]
    }
  }
}

That's the whole config. There is deliberately no API key in it: the first time a tool needs one, ask your assistant to sign in.

Signing in

Say "sign in to EssenceTeacher". The sign_in tool prints a short code and a URL; open it in a browser where you are already signed in, approve the code, and the key is collected once and stored at ~/.config/essenceteacher/api_key (owner-read only). Later sessions just work.

No key is ever pasted into a chat. Only teaching staff can approve — a student account is refused by the server.

If you would rather manage the credential yourself, set ESSENCETEACHER_API_KEY and sign_in is never needed. ESSENCETEACHER_API_URL points at a different deployment; it defaults to https://study.essencescholar.com.

Tools

Getting in — sign_in, whoami

Courses — list_courses, course_details, list_materials, gradebook

Assignments — list_assignments, get_assignment, assignment_usage, list_submissions

Student support — flagged_questions, announcements

Long jobs — list_jobs, get_job

Start at list_courses; its ids feed everything else.

It reads; it does not act

Every tool here is read-only. There is nothing that publishes grades, edits a rubric, posts an announcement or reruns a student's submission — and that is a design decision, not an omission.

Those acts change what a student sees and what a teacher is accountable for. They belong in the app, where the teacher sees the consequence before agreeing to it. An assistant that can publish grades by misreading a sentence is worse than one that cannot publish them at all. When the right next step is one of those, the server's instructions tell the assistant to say so and name the screen.

sign_in is the single exception, and it writes only a credential, with the teacher's own browser consent.

Student data

Rosters, submissions and gradebooks identify real students. Under GDPR the university is the controller, not this tool.

course_details, gradebook and list_submissions say so in their own descriptions, so the assistant knows before it calls them. The server's instructions ask it to answer from aggregates where aggregates will do — a grade distribution rarely needs names attached, and a chat transcript keeps whatever is put into it. Asking for a named list is a teacher's call to make; the point is that names are not reached for by default.

Questions it is good at

• "Which assignment are students struggling with most?" — flagged_questions across courses, grouped by category.
• "Has everyone submitted assignment 34?" — list_submissions, counted.
• "What does this assignment actually require?" — get_assignment for the rubric, in the words the students were given.
• "Is this assignment expensive to run for 80 students?" — assignment_usage.
• "The chatbot keeps saying it does not know about X" — list_materials, to check whether X is in the material set at all.

Licence

MIT.