Canvas Pilot

Website: https://canvas-pilot-rho.vercel.app

Canvas Pilot is an open-source, local-first Canvas LMS AI agent for Codex and Claude Code power users. It learns each course's recurring assignment pattern, then reuses that pattern through scan -> approval -> execute workflows that produce review-ready drafts and run reports.

It is a Canvas MCP workflow layer, not just another Canvas API wrapper. Canvas MCP servers let an agent access Canvas; Canvas Pilot adds recurring course memory, assignment workflow automation, approval boundaries, verification, and student review above the access layer.

The public repo contains the generic framework. Credentials, cookies, private course identifiers, assignment inputs, drafts, and real run outputs stay local and gitignored.

Codex is the sole primary driver. The legacy Claude Code path is preserved as frozen history, but current setup, scan, drafting, submission, quiz, and cron workflows do not depend on it.

Why it is different

See Canvas Pilot vs Canvas MCP for the longer comparison.

Workflow

flowchart LR
    A[Canvas LMS] --> B[scan canvas]
    B --> C[plan.json approval plan]
    C --> D{student approves?}
    D -- no --> E[stop]
    D -- yes --> F[canvas-execute]
    F --> G[course-specific workflow]
    G --> H[result.json per assignment]
    H --> I[REPORT.md closeout]
    I --> J{later exact command?}
    J -- no --> K[human review or manual upload]
    J -- submit N / take quiz N / retake quiz N --> L[signed target-exact receipt]
    L --> M[one Canvas mutation + read-back]

Quick demo

A typical interaction once you're fully configured:

You (in Codex, from this directory):  scan canvas

Canvas Pilot:
    URGENT (next 72h)
    | # | course        | assignment       | due           | proposed skill |
    | 1 | Code Course A | Set 3 Problem 1  | 4/25 23:59    | code-course workflow |
    | 2 | Writing Course| Tue Wk5 HW Scan  | 4/27 12:30    | reading-annotation workflow |
    SOON (3-7d)
    | 3 | Quiz Course   | Section 5        | 5/01 23:59    | quiz workflow |

    Reply: "approve all" / "approve 1, 2" / "urgent only" / "cancel"

You:                                    approve all

Canvas Pilot:
    [dispatches each item sequentially, runs the per-course skill,
     produces a draft under runs/<today>/<assignment>/draft/,
     and writes a status to result.json]

    Done. REPORT.md written at runs/<today>/REPORT.md.
    Drafts mirrored to final_drafts/ for review.

You review the drafts and upload manually by default. Auto-submit only happens through an explicitly configured cron authority. In an interactive session, submission requires a later exact submit N, take quiz N, or retake quiz N message, a verified local result, and a short-lived target-exact receipt.

Public-safe examples live in examples/.

Who this is for

• Students whose school uses Canvas LMS.
• Whose coursework includes a high share of repeating-structure assignments (weekly problem sets, weekly readings, weekly quizzes).
• Who want an AI to handle the boring orchestration (scan, dispatch, draft) but keep the human firmly in the loop for review and submission.
• Who already use Codex, or who are preserving an existing Claude Code workflow.

The default behavior is produce verified drafts, you decide what to submit. Plan approval never grants Canvas write access. Interactive submission is a separate, exact-target action with a verification gate and Canvas read-back.

The product flow

On a new machine, paste the setup prompt from the install page into a local agent. It downloads the repo, asks which school you use Canvas through, opens the official Canvas login, and immediately runs first-skill opportunity analysis after authentication. After that, the routine entry point is scan canvas.

Fresh setup prompt
  1. Download or update the Canvas Pilot repo.
  2. Ask which school you use Canvas through.
  +-- canvas-setup
      3. Resolve the school's official Canvas login URL. If the school answer
         is ambiguous, ask for the Canvas URL instead of guessing.
      4. Prepare local config, open the login browser, and verify authentication.
      5. Immediately hand off to canvas-skill-opportunity when routes are empty.

canvas-skill-opportunity
  1. Discover recurring candidates, then inspect representative real specs.
  2. Judge task fit and safe Canvas feedback-policy evidence qualitatively.
  3. Write a private local recommendation report.
  4. Stop and ask you to choose one numbered candidate.

You choose a candidate in a later reply
  +-- canvas-bootstrap
      1. Verify the real spec, materials, repeatability, and checks.
      2. Scaffold one public-safe course skill and local route.
      3. Keep the route unready if the viability gate fails.

You: "scan canvas" after at least one route is ready
  1. List pending work and write plan.json.
  2. Stop for student approval.
  3. canvas-execute dispatches only approved items.
  4. Each item writes result.json; REPORT.md closes the run.

Three non-obvious things about this flow:

1. Opportunity analysis is read-only and evidence-based. It reads enough representative assignment instructions to understand the real task, plus sanitized facts about retry and feedback policy. It never starts an attempt, solves work, retains raw answers or exact grades, uploads, submits, or creates a route. The agent makes a qualitative recommendation and stops for your numbered choice.

   Strong first-skill defaults include Canvas-submittable code, mostly objective quizzes, quantitative accounting/economics/statistics/math/finance, structured Word/Excel/PDF business work, and independent short answers or annotations. A main continuous prose unit around 200 words or more is a strong demoter; total word count is not the rule. A quiz with multiple attempts, feedback visible before retry, and keep-highest scoring can move forward because it has a useful post-submit correction loop.

2. canvas-bootstrap is single-cluster on purpose. A typical student has 3-4 trackable courses. Configuring all four in one session means a wall of text you can't reason about. Configuring one, walking away, coming back tomorrow for the next gives you a tangible deliverable per session and lets you reuse what you learned from the first cluster when you do the next.

3. The first time you run a per-course skill, it is a co-authoring session with you, not a one-shot. canvas-bootstrap's overlay v1 is the skill's best guess from reading the cluster. v2 is what you actually want, and it only exists after you've reviewed a real draft and your feedback has been categorized. Plan for ~30 minutes per cluster for the first-run loop. Subsequent runs of the same skill are non-interactive.

Setup walkthrough

First time on this machine (you're unconfigured)

1. Copy the concise prompt from https://canvas-pilot-rho.vercel.app/install into a local agent with shell and file access.
2. Let the agent download or update the repo. After download, it asks which school you use Canvas through and resolves that school's official Canvas login URL. It asks for the Canvas URL only when the school answer is ambiguous; it never guesses.
3. Complete login only in the browser it opens. As soon as authentication is verified, the agent runs canvas-skill-opportunity and stops for your numbered choice before route creation, homework scan, execution, or submission.

The prompt dispatches canvas-setup:

• Asks which school you use Canvas through, then resolves its official Canvas login URL. If that cannot be resolved unambiguously, it asks you for the URL instead of guessing.
• Installs Playwright + Chromium silently (~150 MB, one-time).
• Opens a browser. You log in to Canvas with your school SSO once. The browser session stays local. Username/password capture and persistence are off by default; setup removes any legacy record, then verifies that the retained session still authenticates.
• Verifies Canvas authentication without exposing credentials.

canvas-setup then hands off to canvas-skill-opportunity. It discovers recurring candidates, reads enough representative real instructions to judge the task type, considers safe Canvas retry/feedback-policy evidence, writes a private local recommendation, and stops. You reply with one number in a later turn. Only then does canvas-bootstrap verify the real workflow and scaffold your first course skill. Selection does not authorize submission.

Subsequent sessions

You say scan canvas. If courses.yaml.routes has at least one entry but not all your courses, you get the normal scan PLUS a one-line reminder that other courses are still unconfigured. You re-trigger canvas-bootstrap whenever you want to configure another.

Once all your courses are configured, scan canvas is the only command you ever need:

scan canvas
   -> review plan
   -> "approve all" / "approve 1, 2" / "urgent only" / "cancel"
   -> wait for drafts
   -> read REPORT.md, eyeball final_drafts/, upload manually
   -> optionally send one