@canvas-toolchain/shared-llm

canvas-toolchain

Canvas Toolchain helps a professor refresh a Canvas LMS course every semester. It combines three apps — Canvas Toolchain — Curriculum Intelligence, Canvas Toolchain — Design Studio, and Canvas Toolchain — Command & Control — behind a single MCP entrypoint, plus a Python sidecar for downloading Canvas data and a native installer.

Canvas Backup archive
  -> Curriculum Intelligence   (analyze past course, plan the next one)
  -> Canvas Design Studio      (generate Canvas-safe HTML)
  -> Canvas-safe HTML
  -> optional Canvas publishing

Professors drive the whole thing by talking to the Canvas Toolchain MCP server (installed via the native installer, or npx canvas-toolchain if you have Node) from any MCP-capable AI client. Each underlying app also stays independently usable. Direct Canvas API publishing is always optional — the no-token "generate HTML and paste it in" path is first-class.

Optional modules and companion programs: see docs/modules.md.

Where to start

Canvas Toolchain runs as an MCP server you talk to from any MCP-capable AI client.

The installer auto-wires these eight hosts (verbatim from installer/tasks/mcphost.go SupportedHosts()): Claude Desktop, Claude Code, Codex CLI, Gemini CLI, Cursor, VS Code, Kiro, and Antigravity.

Any other MCP-capable client works via the manual JSON snippet below — it is not auto-wired.

Recommended: native installer

Download the Windows x64 / macOS arm64 installer from Releases — it bundles Node, the toolchain, and an auto-updater, and writes the MCP config for every client it detects.

First run: the installer is not code-signed. Windows SmartScreen and macOS Gatekeeper may block it — use More info → Run anyway (Windows) or Open Anyway in System Settings → Privacy & Security (macOS). See the release notes for screenshots.

With Node.js: npx

If you already have Node ≥ 20, you can start the unified MCP server without cloning the repo:

npx canvas-toolchain

That speaks MCP on stdio — wire it into a client below; it is not an interactive CLI.

From source

git clone https://github.com/Ryfter/canvas-toolchain.git
cd canvas-toolchain
npm install        # install IS the build — no separate build step
npx canvas-toolchain   # smoke: starts the MCP server (Ctrl+C to stop)

Requires Node ≥ 20. If your checkout lives in a folder with spaces (e.g. C:\Users\you\Documents\Canvas Toolchain), quote the path anywhere it appears in JSON config.

Wire it into your client

Everywhere below, npx canvas-toolchain also accepts an absolute path form: node <checkout>/packages/command-and-control/dist/index.js.

Claude Desktop (claude_desktop_config.json) / Claude Code (.mcp.json) / Cursor / VS Code — mcpServers JSON:

{
  "mcpServers": {
    "canvas-toolchain": { "command": "npx", "args": ["canvas-toolchain"] }
  }
}

Codex CLI:

codex mcp add canvas-toolchain -- npx canvas-toolchain

Gemini CLI (~/.gemini/settings.json):

{ "mcpServers": { "canvas-toolchain": { "command": "npx", "args": ["canvas-toolchain"] } } }

Any other MCP-capable client (not in the installer host table above): any client that can run a stdio MCP server works via this same JSON snippet — command npx, args ["canvas-toolchain"]. Restart the client after editing its config. That includes hosts the installer does not auto-wire.

Migration: existing configs that point at the old bin names (canvas-design-mcp, command-and-control-mcp) should switch to npx canvas-toolchain (or the new bins canvas-toolchain-server, canvas-toolchain-design-studio, canvas-toolchain-curriculum-intelligence if you still want a single-package server).

Working on this repo (human or AI agent)? Read AGENTS.md first.

Documentation

Doc	What's in it
Feature overview	Concise outcome-focused summary of what Canvas Toolchain helps instructors do, with emphasis on accessibility, Canvas management, shell speed/quality, and instructor-led AI review
Roadmap	What ships next and in what order — the WCAG 2.2 gate phases, upcoming releases, the v2.0 plug-in direction, and the unscheduled ideas backlog
User Guide & Tutorial	Start here. What you can do, how to use it, a hands-on end-to-end tutorial, and a task-by-task command catalog (what each command is · how it works · why you'd use it)
Commands & Credentials reference	Every command (MCP tool) with its parameters, and every API key/secret — what it's for, why, and whether it's optional
Module view	The toolchain broken into modules: what each is, why it exists, what it does, and its commands
Accessibility checks	Every automated accessibility check the toolchain runs — what each one catches, how it works (with the contrast math), where it runs (advisory vs blocking), how to fix findings, and its limitations
Visual guide	Picture-first tour — diagrams of the pipeline, architecture, and setup (rendered PNG/SVG + an editable Excalidraw scene)

What lives where

Path	Package	What it owns
packages/canvas-toolchain/	canvas-toolchain	npx entrypoint launcher — the published bin professors run (npx canvas-toolchain)
packages/command-and-control/	@canvas-toolchain/command-and-control	Single professor-facing MCP server (canvas-toolchain-server); workflow orchestration, registry, adapters
packages/canvas-design-studio/	@canvas-toolchain/canvas-design-studio	Canvas-safe HTML generation (canvas-toolchain-design-studio), design review, transcript enrichment
packages/curriculum-intelligence/	@canvas-toolchain/curriculum-intelligence	Course analysis, semester comparison, topic currency, planning (canvas-toolchain-curriculum-intelligence)
packages/shared-types/	@canvas-toolchain/shared-types	TypeScript contracts shared across packages
packages/shared-llm/	@canvas-toolchain/shared-llm	Shared LLM client (Anthropic + Ollama providers)
packages/module-contract/	@canvas-toolchain/module-contract	The CanvasToolchainModule plug-in contract for opt-in capability modules
packages/module-video/	@canvas-toolchain/module-video	Lecture Video module (Panopto as the first provider)
packages/module-oral-assessment/	@canvas-toolchain/module-oral-assessment	Oral/video assessment authoring (Rhetorix-first)
packages/module-group-builder/	@canvas-toolchain/module-group-builder	Create and rotate balanced student groups
packages/module-roster/	@canvas-toolchain/module-roster	Roster & Identity Manager (PeopleSoft → de-identified roster + pseudonyms)
packages/module-peerassessment/	@canvas-toolchain/module-peerassessment	Export a Canvas group set to a PeerAssessment.com import CSV
installer/	—	Go + Fyne native installer and auto-updater
canvas-backup (separate repo)	—	Python Canvas backup downloader, reached via a CLI bridge

Verification

From the repo root:

npm test          # all TypeScript workspace tests
npm run build     # compile all packages
npm run smoke:integration --workspace=packages/command-and-control

Installer (from installer/):

go vet ./...
go test ./...
go build -o ct-installer.exe .

See AGENTS.md for the full cross-package verification matrix and the installer's local-payload build notes.