Skip to content

Agent Skill

Canvas CLI ships an agent skill — a SKILL.md (plus condensed reference files) that teaches AI coding agents (Claude Code, Cursor, Codex, Gemini CLI, Windsurf, GitHub Copilot, and more) how and when to drive the canvas CLI. Once installed, you can just ask your agent things like "grade everyone in grades.csv for course 123" or "create next week's assignments" and it knows the commands, flags, and safety rules.

Install the skill

The cross-agent way (recommended) — detects every agent you have and stays up to date:

npx skills add jjuanrivvera/canvas-cli

Built into the CLI (no Node required) — writes the bundled skill directly:

canvas skills install                 # current project (./.claude/skills)
canvas skills install --global        # user-level (~/.claude/skills)
canvas skills install --agent cursor --global
canvas skills path                    # show where it would install
canvas skills print                   # print the SKILL.md

Supported --agent values: claude, cursor, windsurf, codex, gemini, copilot, opencode. Use --dir to write to any custom skills directory.

Native Claude Code plugin:

/plugin marketplace add jjuanrivvera/canvas-cli
/plugin install canvas-cli@canvas

Prerequisites

The skill wraps the canvas binary — it doesn't bundle it. So:

  1. Install the CLI: brew tap jjuanrivvera/canvas-cli && brew install canvas-cli (or go install github.com/jjuanrivvera/canvas-cli/cmd/canvas@latest).
  2. Authenticate: canvas auth login or canvas auth token set (or set CANVAS_URL/CANVAS_TOKEN).
  3. Verify: canvas doctor.

Skill vs MCP

  • Skill (this page) — markdown instructions for shell-capable agents. The agent runs canvas … commands directly. Lowest friction; best for CLI use.
  • MCP server — canvas mcp start exposes every command as a structured tool, for clients that prefer tool schemas (Claude Desktop, etc.).

Use whichever your agent supports; they can coexist.

What the agent learns

The skill teaches the auth → discover → act → verify workflow, the golden rules (preview writes with --dry-run, parse with -o json, resolve IDs live, bound large lists with --limit, confirm destructive actions), the resource/action map, output filtering (--filter/--columns/--sort), grading workflows (single and bulk CSV), course-content publishing, cross-instance sync, the raw canvas api escape hatch, and error handling.

Condensed references ship alongside it:

  • references/canvas-commands.md — command/flag cheatsheet
  • references/auth-and-config.md — auth methods, env vars, instances, context
  • references/output-and-filtering.md — formats, jq patterns, script hygiene