Canvas MCP

Canvas MCP puts a Canvas LMS instance behind the Model Context Protocol, and it is unusually thorough about it: up to 101 tools, organised by who is asking rather than by which API endpoint they map to.

Tools grouped by role

• Students — assignment tracking, grades, peer reviews, submissions.
• Educators — assignment management, grading, course analytics, messaging.
• Learning designers — accessibility scanning and course structure analysis.
• Shared — courses, discussions, pages and modules.
• Developers — code execution and tool discovery.

The accessibility auditing and course-structure analysis are the parts that are hard to get any other way; scanning a whole course for accessibility problems is exactly the sort of tedious, rule-driven review an agent is better at than a human with a deadline.

Agent skills too

The project ships eight companion agent skills that encode common workflows rather than single calls: canvas-week-plan, canvas-morning-check, canvas-bulk-grading, canvas-peer-review-manager, canvas-discussion-facilitator, canvas-course-qc, canvas-accessibility-auditor and canvas-course-builder.

Installing it

The quickest route for any agent is npx skills add vishalsachdev/canvas-mcp. For Claude Desktop specifically, download the canvas-mcp.mcpb bundle from the releases page and double-click it. A manual Python install (virtualenv, pip install -e ., and a .env from the provided template) is available for anyone who wants to run it from source.

This is an independent community project, MIT licensed, not an Instructure product — you will need a Canvas API token from your own institution's instance, and whatever your institution permits that token to do is the real limit on what the server can do.