CanvasPilot

Agentic Canvas LMS tools for LLMs — an MCP server + CLI that talks to Canvas's real REST API using the session your browser already has.

Most Canvas MCPs need a Personal Access Token. Many schools disable student PATs and hide Canvas behind SSO/MFA. CanvasPilot solves that with a stay-open session broker: you log in once in a real browser (Playwright, persistent profile), then every tool call rides that authenticated session — headless, for as long as the cookies live.

Same architecture as OpenCLI-style web agents (persistent browser session → site's real HTTP/JSON API), purpose-built for Canvas.

Features

• Auth that works at real schools — session broker (SSO/MFA, no PAT needed) or classic CANVAS_API_TOKEN

• Headless after one headed login — session start --headless reuses the saved profile

• Full REST surface — courses, assignments, modules, pages, files, discussions, announcements, planner, inbox, calendar, activity stream, submissions, and quizzes (list/questions/submissions/start/complete)

• Agent-shaped digests — assignment_brief (cleaned prompt + rubric), sync_summary (courses + upcoming), submission_status

• Fixture mode — offline dict backend for tests and CI; no Canvas required

• 28 MCP tools, one stdio server, zero config beyond CANVAS_BASE_URL

Related MCP server: Canvas Assignment Assistant

Install

pip install "canvaspilot[browser]"      # or: pip install -e ".[browser]" from a checkout
playwright install chromium

Requires Python 3.11+.

Quick start

export CANVAS_BASE_URL=https://<school>.instructure.com

# 1) Log in once (headed — finish SSO/MFA in the window), leave the broker running
canvaspilot session start

# 2) Later: restart headless on the same saved profile
canvaspilot session start --headless

# 3) Use it
canvaspilot whoami
canvaspilot courses
canvaspilot sync
canvaspilot brief <course_id> <assignment_id>

With a PAT instead:

export CANVAS_API_TOKEN=...   # broker not needed
canvaspilot courses

MCP server

canvaspilot mcp          # stdio

Cursor / Claude Desktop / any MCP client:

{
  "mcpServers": {
    "canvas": {
      "command": "canvaspilot",
      "args": ["mcp"],
      "env": { "CANVAS_BASE_URL": "https://<school>.instructure.com" }
    }
  }
}

Tools exposed (all prefixed canvas_):

Auth modes

The broker only listens on loopback. Cookies never leave the Playwright profile directory; the MCP/CLI process never sees them — it asks the broker to make the request.

Configuration

Session broker commands

canvaspilot session start [--headless] [--base-url URL] [--profile DIR] [--port N]
canvaspilot session status
canvaspilot session stop

Scripts

• scripts/readonly_sweep.py — live read-only smoke across the whole tool surface (needs a running broker)

• scripts/inventory_pass.py — per-course inventory: tabs, LTI/external tools, modules, third-party mentions

Development

pip install -e ".[dev]"
pytest

Tests run entirely in fixture mode.

Responsible use

CanvasPilot gives an agent the same access you have — including submitting assignments and starting quiz attempts. Write tools are clearly named; wire approval gates in your agent if you don't want autonomous submits. Follow your institution's academic integrity policy.

Origin

Extracted from the Suite monorepo (packages/canvaspilot), where it's synced via git subtree.

License

MIT