Canvas Reader

A lean, fast, read-only way for AI assistants to read everything in a student's own Canvas: deadlines, assignments, grades, announcements, the Inbox, syllabi, pages, modules, discussions and course files. It runs on your Mac with your own Canvas token and works with Claude, ChatGPT and Codex.

Claude Code, Claude Desktop, Codex ───────────────────────► Canvas Reader on your Mac ─► your Canvas
ChatGPT (web, desktop, phone), Work ─► Secure MCP Tunnel ─┘

Canvas Reader builds on canvas-mcp 1.13.0 (pinned). What's different:

Student-facing and read-only, always. It exposes only canvas-mcp's student read tools. Every tool that submits, edits, sends messages or runs code is removed at startup, whatever the settings say.
Complete, or it says so. When Canvas or a file fails partway, the answer is marked PARTIAL RESULT and says what's missing. It never presents part of your assignments as all of them.
Fast and lean. "What's due?" across every course is one call. Long files are read in chunks and kept in memory for 15 minutes, so follow-up questions don't download them again. Instructions and output are kept compact to save your assistant's context.
Non-invasive. Nothing is written to disk, there's no telemetry, and the assistant is told to ask before opening Canvas in a browser.

What it reads

30 read-only tools: canvas-mcp's student read tools (courses, assignments and your submissions with feedback, grades, announcements, discussions, Inbox conversations, pages, modules, files, to-do items, peer reviews, your profile) plus four of its own. Upstream tools that only repeat these less accurately are left out, which keeps the list your assistant reads short. Canvas Reader's own tools:

Tool	What it does
`get_upcoming_deadlines`	What's due in the next N days (default 7) across all active courses, plus recent overdue work, with local times, status and links, in one call.
`get_course_assignment_data`	One course's full assignment list, including undated work, with your own due dates and submission status.
`read_course_document`	Text of a PDF, Word, PowerPoint, HTML, RTF or text file (up to 25 MiB), in chunks with page or slide markers.
`search_course_document`	Find a phrase in a course file and jump to it.

Word files include content controls, tables, text boxes, headers, footers, footnotes, comments, charts and SmartArt. PowerPoint files include tables, charts, speaker notes, comments and hidden slides. Images aren't read (no OCR); they're counted in a warning, as are scanned PDF pages.

Try: "What's due this week?", "Summarize this week's lecture slides in Biology", "What does the syllabus say about late work?", "Any new announcements?"

Privacy and your data

Your Canvas token stays on your Mac: in a private credential file, your macOS Keychain, or Claude's secure storage. It never goes into the release files or tool output.
Downloaded course text stays in memory, expires after 15 minutes and is wiped when the reader stops. Nothing is written to disk. Logs hold warnings and errors, not course text.
What a tool returns (course content, Inbox messages, grades) goes to the assistant and model provider you use. Check their data policies.
Course content is labeled as untrusted, and the assistant is told to treat it as data, never as instructions.

SECURITY.md has the full threat model and how to report a problem.

Works with

Client	Support	Guide
Claude Code (terminal, IDE, desktop app)	Yes: one command, or the plugin	Claude
Claude Desktop	Yes: one-click extension, or the config file	Claude
Claude cloud tasks linked to your Mac	Should work through Claude Desktop; verify	Claude
ChatGPT on the web, desktop and phone	Yes, through OpenAI's Secure MCP Tunnel	ChatGPT
ChatGPT Work	Yes, through the tunnel	ChatGPT
Codex in the ChatGPT desktop app	Yes, through the tunnel or directly	ChatGPT, Codex
dot	Not confirmed; may work through the tunnel	ChatGPT
Codex CLI and IDE extension	Yes: one command	Codex
claude.ai on the web and phone	No: custom connectors reach only remote HTTPS servers	Why
Claude Code on the web	No: it would need your token in the cloud	Why
Codex cloud	No: MCP isn't available in cloud tasks (the skill is)	Why

Quick start

You need a Mac, Python 3.11 or newer (python3 --version; get it from python.org or brew install python), git, and a Canvas access token (Canvas → Account → Settings → New access token).

Get the code and install the runtime. This creates its own Python environment under ~/.local/share/canvas-reader, checks that it starts, and makes it active.
```
git clone https://github.com/NeerAeron/canvas-reader.git ~/canvas-reader
cd ~/canvas-reader
python3 scripts/versions.py install .
```
Add your Canvas credentials to ~/.config/canvas-reader/canvas.env: SETUP.md step 2.
Check it against your Canvas account: SETUP.md step 3.
Connect your assistant: Claude, ChatGPT or Codex.

docs/SETUP.md is the full Mac guide, including installing from a downloaded release instead of git.

Updating and removing

cd ~/canvas-reader
git pull
python3 scripts/versions.py update     # installs the newer version, checks it, switches to it
python3 scripts/versions.py rollback   # switches back if something is wrong

Without git, python3 scripts/versions.py update --github downloads the newest release and checks it against its published checksums. Restart your assistant (or the tunnel) after updating.

To remove everything, see SETUP.md → Remove everything.

Contributing

See CONTRIBUTING.md. Report security problems privately as described in SECURITY.md.

License and disclaimer

Not affiliated with or endorsed by Instructure, OpenAI or Anthropic. Canvas is a trademark of Instructure.