Canvas MCP

MCP (Model Context Protocol) server for Canvas LMS. Exposes Canvas courses, assignments, modules, pages, files, grades, and submissions as tools for AI agents.

Tools

Related MCP server: Canvas MCP

Resources

Setup

Environment Variables

export CANVAS_API_TOKEN="your-canvas-token"
export CANVAS_BASE_URL="https://your-institution.instructure.com"

Generate a token at: <your-canvas-url>/profile/settings → + New Access Token

Install

pip install -e .

Run

canvas-mcp              # stdio transport (for MCP clients)

Add to Hermes config.yaml

mcp_servers:
  canvas:
    command: canvas-mcp
    env:
      CANVAS_API_TOKEN: "your-token"
      CANVAS_BASE_URL: "https://psu.instructure.com"

Add to Claude Desktop

{
  "mcpServers": {
    "canvas": {
      "command": "canvas-mcp",
      "env": {
        "CANVAS_API_TOKEN": "your-token",
        "CANVAS_BASE_URL": "https://psu.instructure.com"
      }
    }
  }
}

Design Decisions

• No stdout output from tools — tools return structured dicts/lists only, safe for coding agents

• HTML is stripped — all Canvas HTML content is converted to readable plain text

• Responses are slimmed — only relevant fields returned, no noise

• Pagination handled — follows Canvas Link headers automatically (up to 10 pages)

• Graceful 403 handling — list_files returns an error dict with a hint instead of crashing

• Client cleanup — every tool call opens and closes its own client connection

Testing

pip install pytest pytest-asyncio
pytest -v

Tests use a MockTransport that intercepts HTTP at the httpx layer — no real Canvas calls. Three test files:

• test_client.py — Canvas API client unit tests

• test_server.py — tool functions, HTML stripping, slimming helpers

• test_integration.py — full pipeline: tool → client → mock HTTP → structured response

License

MIT