@krishpkreame/canvas-mcp-server

v2.5.1

Published

12 days ago

A read-only Model Context Protocol (MCP) server for Canvas LMS — safe for student use with no state-mutating operations

0High
0Medium
0Low

krishpkreame

canvas lms mcp server modelcontextprotocol education api student learning assignments courses

Canvas MCP Server v2.5.0

Forked from DMontgomery40/mcp-canvas-lms
This fork strips the server down to 26 read-only tools for safe student use. All state-mutating operations have been removed — no assignments can be submitted, no profiles updated, no module items marked complete. Nothing in Canvas is ever modified.
Why this fork? I'm a student at the University of Auckland (UoA) and wanted a Canvas MCP I could safely connect to Claude without any risk of accidental writes to my account. The upstream repo is a full-featured server for students, instructors, and admins — this fork is just the read side.

A read-only Model Context Protocol (MCP) server for Canvas LMS — safe for student use with no state-mutating operations

🎯 Key Features

🎓 For Students

Course Management: Access all courses, syllabi, and course materials
Assignment Tracking: View assignments, check submission status, and track due dates
Announcements: Read course announcements
Progress Tracking: Monitor grades, module completion, and calendar events
Quizzes: View quizzes and quiz details
File Access: Browse and download course files and resources

🛠️ Technical Excellence

Read-Only by Design: No mutating operations — safe to connect to any Canvas account
In-Memory Caching: TTL-based response cache with LRU eviction (10m / 5m / 1m tiers) — up to 40x faster on repeated calls
Request Deduplication: Concurrent calls for the same endpoint share a single HTTP request
Token Efficient: Compact JSON serialization and lean tool descriptions save ~30-40% tokens in LLM context
Robust API: Automatic retries, per_page=100 pagination (10x fewer round-trips), comprehensive error handling
Type Safe: Full TypeScript implementation with strict types
26 Tools: Focused, read-only coverage of Canvas LMS

Quick Start

Option 1: Clone and run locally (recommended)

git clone https://github.com/KrishPatel1404/mcp-canvas-lms.git
cd mcp-canvas-lms
npm install && npm run build

Then add to your MCP client config (Claude Desktop, opencode, Cursor, etc.):

{
  "mcpServers": {
    "canvas-lms": {
      "command": "node",
      "args": ["/path/to/mcp-canvas-lms/build/index.js"],
      "env": {
        "CANVAS_API_TOKEN": "your_token_here",
        "CANVAS_DOMAIN": "your_school.instructure.com"
      }
    }
  }
}

Option 2: opencode

Add to ~/.config/opencode/opencode.json (or a project-level opencode.json):

{
  "mcp": {
    "canvas-lms": {
      "type": "local",
      "command": ["node", "/path/to/mcp-canvas-lms/build/index.js"],
      "environment": {
        "CANVAS_API_TOKEN": "your_token_here",
        "CANVAS_DOMAIN": "your_school.instructure.com"
      }
    }
  }
}

Option 3: NPM (once published)

npm install -g @krishpkreame/canvas-mcp-server

{
  "command": "canvas-mcp-server",
  "env": {
    "CANVAS_API_TOKEN": "your_token_here",
    "CANVAS_DOMAIN": "your_school.instructure.com"
  }
}

Transport Modes

The server supports two explicit transport modes:

stdio (default): best for Claude Desktop/Codex/Cursor local MCP wiring.
streamable-http: best for local HTTP integrations and containerized workflows.

Transport environment variables

# Required Canvas auth
CANVAS_API_TOKEN=your_token
CANVAS_DOMAIN=your_school.instructure.com

# Transport selection
MCP_TRANSPORT=stdio # or streamable-http

# Streamable HTTP settings
MCP_HTTP_HOST=127.0.0.1
MCP_HTTP_PORT=3000
MCP_HTTP_PATH=/mcp
MCP_HTTP_STATEFUL=true
MCP_HTTP_JSON_RESPONSE=true
MCP_HTTP_ALLOWED_ORIGINS=

Performance

All optimizations are built-in with zero configuration required.

| Optimization | Impact | |---|---| | Response cache (TTL: 10m / 5m / 1m) | Up to 40x latency reduction on repeated calls | | per_page=100 default | Up to 10x fewer HTTP round-trips for paginated endpoints | | Request deduplication | Concurrent calls for the same data share a single request | | Compact JSON | ~30-40% fewer tokens in LLM context window | | Lean tool descriptions | ~500+ tokens saved per tools/list response |

Cache TTL Tiers

| Tier | TTL | Endpoints | |---|---|---| | Long | 10 min | User profile, course list, dashboard cards, syllabus | | Medium | 5 min | Assignments, modules, pages, files, folders, quizzes | | Short | 1 min | Submissions, grades, calendar, notifications, announcements |

Tuning

CANVAS_TIMEOUT=60000      # Request timeout in ms (default: 30000)
CANVAS_MAX_RETRIES=3      # Retry count for 429/5xx errors
CANVAS_RETRY_DELAY=1000   # Base retry delay in ms (exponential backoff)

🎓 Workflow Examples

Check Today's Assignments

"What assignments do I have due this week?"

Lists upcoming assignments with due dates, points, and submission status

Check Grades

"What's my current grade in Biology?"

Shows current scores, grades, and assignment feedback

Read Announcements

"Any new announcements in my courses?"

Lists recent course announcements

Track Progress

"What modules do I need to complete in Math 200?"

Shows module completion status and next items to complete

Getting Canvas API Token

Log into Canvas → Account → Settings
Scroll to "Approved Integrations"
Click "+ New Access Token"
Enter description: "Claude MCP Integration"
Copy the generated token Save securely!

⚠️ A standard student token is all that's needed — no admin privileges required.

Development

git clone https://github.com/KrishPatel1404/mcp-canvas-lms.git
cd mcp-canvas-lms
npm install

# Start development with hot reload
npm run dev:watch

# Code quality
npm run lint
npm run type-check

📚 Available Tools — 26 read-only tools

All tools are read-only. Nothing in Canvas is ever modified.

System

| Tool | Description | |------|-------------| | canvas_health_check | Check API connectivity and confirm your token works |

Courses

| Tool | Description | |------|-------------| | canvas_list_courses | List all enrolled courses | | canvas_get_course | Get detailed info for a specific course | | canvas_get_syllabus | Get the syllabus body for a course | | canvas_get_dashboard_cards | Get dashboard course cards |

Assignments & Submissions

| Tool | Description | |------|-------------| | canvas_list_assignments | List all assignments in a course | | canvas_get_assignment | Get details for a specific assignment | | canvas_list_assignment_groups | List assignment groups and weights | | canvas_get_submission | View your submission for an assignment | | canvas_get_upcoming_assignments | Get upcoming due dates across all courses |

Grades & Calendar

| Tool | Description | |------|-------------| | canvas_get_course_grades | Get your current grade and scores for a course | | canvas_list_calendar_events | List calendar events within a date range |

Modules

| Tool | Description | |------|-------------| | canvas_list_modules | List all modules in a course | | canvas_get_module | Get details for a specific module | | canvas_list_module_items | List all items inside a module | | canvas_get_module_item | Get details for a specific module item |

Announcements & Notifications

| Tool | Description | |------|-------------| | canvas_list_announcements | List course announcements | | canvas_list_notifications | List your activity stream notifications |

Files & Pages

| Tool | Description | |------|-------------| | canvas_list_files | List files in a course or folder | | canvas_get_file | Get details and download URL for a file | | canvas_list_folders | List folders in a course | | canvas_list_pages | List wiki pages in a course | | canvas_get_page | Get the content of a specific page |

Quizzes

| Tool | Description | |------|-------------| | canvas_list_quizzes | List all quizzes in a course | | canvas_get_quiz | Get details for a specific quiz |

User

| Tool | Description | |------|-------------| | canvas_get_user_profile | Get your Canvas profile information |

🌟 Example Claude Conversations

Student: "What assignments do I have due this week?"

Claude: Let me check your upcoming assignments...

[Claude uses canvas_get_upcoming_assignments]

Student: "What's my current grade in Biology?"

Claude: Let me pull your grades for that course...

[Claude uses canvas_get_course_grades]

Student: "Show me the modules I need to complete in Math 200"

Claude: I'll list the modules and their completion status for you...

[Claude uses canvas_list_modules, then canvas_list_module_items]

🔍 Troubleshooting

Common Issues:

❌ 401 Unauthorized: Check your API token is correct and not expired
❌ 404 Not Found: Verify course/assignment IDs and that you're enrolled in the course
❌ 403 Forbidden: Some endpoints (e.g. rubrics) require instructor access — expected for students
❌ Timeout: Increase CANVAS_TIMEOUT in your .env or check network connectivity

Debug Mode:

export LOG_LEVEL=debug
npm start

🤝 Contributing

Quick Contribution Setup

git clone https://github.com/KrishPatel1404/mcp-canvas-lms.git
cd mcp-canvas-lms
npm install
npm run dev:watch
# Make changes, add tests, submit PR

🙋 Support & Community

🐛 Bug Reports: GitHub Issues
💬 Questions: GitHub Discussions
📦 Upstream: DMontgomery40/mcp-canvas-lms

Appendix: MCP in Practice (Code Execution, Tool Scale, and Safety)

Last updated: 2026-03-16

Why This Appendix Exists

MCP is still one of the most useful interoperability layers for agentic tooling. The tradeoff is that large MCP servers can expose dozens of tools, and naive tool-calling can flood context windows with tool schemas, call traces, and low-signal chatter.

In practice, larger tool surfaces only help when orchestration stays token-efficient and execution behavior is constrained.

The Shift to Code Execution / Code Mode

Recent production workflows move orchestration out of conversational turns and into executable loops. This keeps context overhead lower, improves determinism, and makes runs auditable.

Core reading:

Recommended Setup for Power Users

For lower-noise, repeatable MCP usage, start with codemode-oriented routing:

Even with strong setup, model behavior can be hit-or-miss across providers and versions. Keep retries and deterministic fallbacks.

Peter Steinberger Workflow Pattern

A high-leverage pattern is turning broad MCP tool surfaces into narrower CLI/task interfaces:

What Works Best With Which MCP Clients

Claude Code / Codex / Cursor agent workflows: usually strong for direct MCP + code-execution loops.
Thin hosted chat clients: often safer with wrapped CLIs/gateways instead of full raw tool exposure.
High-tool-count servers: usually better when split into narrow task gateways.

This ecosystem changes quickly. If you are reading this now, parts of this section may already be out of date.

Prompt Injection: Risks, Consequences, and Mitigations

Prompt injection remains an open problem for tool-using agents. It is manageable, but not solved.

Primary risks:

Hidden instructions in retrieved content or tool output.
Secret/token exfiltration through unintended calls.
Unauthorized state changes in systems or data.

Mitigation baseline:

Least-privilege credentials and scoped tokens.
Destination/action allowlists and strict schema validation.
Human confirmation for destructive operations.
Sandboxed execution and resource limits.
Structured logging and replayable execution traces.

Treat every tool output as untrusted input unless explicitly verified.

📄 License

MIT License - see LICENSE file for details.

⭐ Star this repo if it helps you! ⭐