PokeCanvas
Enables querying Canvas LMS courses, assignments, grades, and more via natural language through Poke. Reads live data on every request.
README
Canvas LMS integration for Poke
Connect your Canvas LMS account to Poke by The Interaction Company so you can ask Poke about your classes in plain language:
"What's due this week?" · "Did my CS101 grade change?" · "Read me the syllabus page" · "Any new announcements?"
It's a small FastMCP server that wraps the Canvas REST API and speaks the Model Context Protocol (MCP) Poke connects to. Once deployed, Poke discovers the tools automatically and uses them whenever a question is about your coursework.
All data is read live on every request, so answers always reflect the current state of Canvas — new assignments, changed grades, fresh announcements, updated due dates. There is nothing to re-sync.
⚡ 1-click setup
Hosted — no terminal. Click, set 2 fields, deploy:
On the deploy screen set CANVAS_BASE_URL + CANVAS_API_TOKEN. Then open your new service URL in a browser — the page shows your live MCP link and a one-tap-copy npx poke@latest mcp add … command already filled in with your URL. Paste it and Poke is connected. (Set the optional POKE_API_KEY too and the server texts you a "Canvas is live" confirmation on boot.)
Local — one paste:
curl -fsSL https://raw.githubusercontent.com/jackulau/PokeCanvas/main/bootstrap.sh | bash
This bootstrap.sh installs everything, asks for your Canvas URL + token once, opens a public tunnel, and registers it with Poke for you.
Other hosts (Railway, Fly, Docker, VPS) → HOSTING.md. Step-by-step detail below.
How it's hosted — two models
This is built as a multi-tenant server, so there are two ways to run it:
- Shared instance (least setup): one deployment serves many people. Each user adds the shared
/mcpURL in Poke and passes their own Canvas credentials as the integration key —https://canvas.yourschool.edu::YOUR_CANVAS_TOKEN. No per-user hosting, nothing to deploy. The server is hardened for untrusted callers: SSRF protection on the Canvas URL, per-user rate limiting (X-Poke-User-Id), and it never logs or stores tokens. See SECURITY.md. - Self-host (most private): deploy your own (the 1-click above), set
CANVAS_BASE_URL+CANVAS_API_TOKEN, connect with no key. Your token never leaves infrastructure you control.
Not sure which? Self-host — it's one click and the token stays yours.
What Poke can see
Every resource you asked for, each as an MCP tool Poke calls on demand:
| Area | Tools |
|---|---|
| Courses | list_courses (with current grade + term), get_profile |
| Assignments | list_assignments (incl. your score / submission status / late / missing) |
| Files | list_files |
| Pages | list_pages, get_page (full content) |
| Modules | list_modules (with the items inside each) |
| Announcements | list_announcements (one course, or all at once) |
| Discussions | list_discussions |
| Quizzes | list_quizzes |
| Calendar | list_calendar_events, list_upcoming_events |
| To-dos / changes | list_todos, get_recent_activity (the live "what changed" feed) |
Read-only by design — it never submits, posts, or deletes anything in Canvas.
Setup
Three steps. Fastest path with no signup: ./setup.sh — it installs deps, asks for
your Canvas URL + token once, opens a public tunnel, and registers it with Poke for you.
Prefer real 24/7 hosting? Full per-host instructions live in HOSTING.md.
1. Get your Canvas access token
In Canvas: Account → Settings → Approved Integrations → "+ New Access Token". Copy it (shown once). Note your Canvas URL — where you log in, e.g. https://canvas.university.edu.
2. Deploy the server — pick one
One-click hosted:
Or local, no signup:
git clone https://github.com/jackulau/PokeCanvas && cd PokeCanvas
./setup.sh
Set CANVAS_BASE_URL + CANVAS_API_TOKEN on whichever host you pick. Every option ends with a URL like https://<host>/mcp — note the /mcp, no trailing slash. Render / Railway / Fly / Heroku / Docker / VPS / tunnel details → HOSTING.md.
Render's free tier sleeps after ~15 min idle (first request after a nap ~50s). Railway/Fly/Heroku or your own Docker host avoid cold starts.
3. Connect it to Poke
Because the credentials live on your server, Poke needs no API key — just the URL.
- One command:
npx poke@latest mcp add https://<host>/mcp -n "Canvas" - Or web UI: poke.com/settings/connections → New integration → Name
Canvas, URLhttps://<host>/mcp, leave the API key blank → Create. - Or shareable recipe (Poke sets it up for any user from one link):
recipe/recipe.md.
Then message Poke "What courses am I taking?"
"Dynamically update if anything changes"
There's no stale cache to worry about: every tool call hits Canvas in real time. Three things make change-tracking explicit:
get_recent_activity— Canvas's activity stream: new announcements, grade changes, submission comments, discussion replies.list_todos/list_upcoming_events— what needs action and what's due, soonest first.list_assignmentsreturns your live submission state (graded? late? missing? score?).
Want Poke to proactively ping you when something changes (not just when you ask)? Add a scheduled automation — the recipe in recipe/ includes a ready-to-use prompt for it ("every morning, check get_recent_activity and text me anything new").
Auth modes
| Mode | When | How |
|---|---|---|
| Env vars (default) | You run your own server | Set CANVAS_BASE_URL + CANVAS_API_TOKEN on the host. Connect to Poke with no key. |
| Bearer token | One server, you pass the token from Poke | Put your token in Poke's API-key field. Base URL still comes from CANVAS_BASE_URL (or send an X-Canvas-Base-Url header). |
| Composite bearer | A shared server for many schools (e.g. a recipe) | Poke API key = https://canvas.yourschool.edu::YOUR_TOKEN. The server splits it per user. |
Precedence: a token in the request beats the env token; the request base URL (composite or header) beats CANVAS_BASE_URL.
Local development
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements-dev.txt
# Run it
export CANVAS_BASE_URL=https://canvas.youruni.edu
export CANVAS_API_TOKEN=your_token
python src/server.py # serves http://localhost:8000/mcp
# Verify (no Canvas account needed — uses mocks + a live protocol check)
pytest -q # unit tests
python scripts/smoke_tools_list.py # all tools registered -> TOOLS_OK
ruff check src tests scripts # lint
To test the live transport the way Poke sees it:
PORT=8765 python src/server.py &
python scripts/live_http_test.py http://127.0.0.1:8765/mcp # -> LIVE_OK
Troubleshooting
| Symptom | Fix |
|---|---|
| First request hangs ~50s | Render free-tier cold start; it's awake now. Use a paid/always-on host to avoid. |
401 ... access token |
Token wrong/expired, or CANVAS_API_TOKEN not set on the host. Regenerate in Canvas. |
401 ... base URL |
CANVAS_BASE_URL not set, or wrong (must be your login URL, e.g. https://canvas.uni.edu). |
| Poke "can't reach the tool" | Confirm the URL ends in /mcp (no trailing slash). Try clearhistory in Poke, then re-test. |
| Empty course list | The token's user may have no active enrollments; try "list all my courses including past ones". |
How it works
Poke ──(MCP over HTTPS, streamable-http at /mcp)──► this server ──(REST + Bearer)──► Canvas LMS API
src/server.py— FastMCP app (transport="http",stateless_http=True), tool + route registration, landing page (/), health (/health), optional boot ping.src/tools.py— the 14@mcp.tooldefinitions; rate-limit, build a client from the request, call the data layer, return clean errors.src/canvas_api.py— one function per resource; maps to Canvas endpoints and trims responses to what matters.src/canvas_client.py— async HTTP, credential resolution, and transparent CanvasLink-header pagination.src/security.py— SSRF guard for the caller-supplied Canvas URL (multi-tenant).src/ratelimit.py— per-user sliding-window rate limiter.src/landing.py— the self-configuring connect page served at/.
See recipe/recipe.md for the Poke recipe and onboarding copy.
Security
Read-only, stateless, tokens never stored or logged, HTTPS-only, with SSRF protection and per-user rate limiting for shared deployments. Full threat model and the self-host-vs-shared tradeoff: SECURITY.md. Report vulnerabilities via a private GitHub security advisory.
Contributing
Issues and PRs welcome.
git clone https://github.com/jackulau/PokeCanvas && cd PokeCanvas
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements-dev.txt
pytest -q && ruff check src tests scripts # keep both green
Please add a test for any new tool or behavior and keep ruff clean.
License
MIT © 2026 jacklau
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。