I built a local-first Canvas LMS agent workflow for Codex and Claude Code users

#ai #programming #opensource #productivity

I have been building Canvas Pilot, a local-first Canvas LMS AI agent for recurring coursework workflows.

The main idea is not "another Canvas API wrapper." Canvas MCP servers and Canvas API clients can help an agent see Canvas. Canvas Pilot is the workflow layer above that:

scan Canvas assignments into an approval plan
stop before execution
let the student approve selected work
route approved items into recurring course workflows
produce review-ready drafts, result files, and REPORT.md
keep private course data and drafts local

Why workflow memory matters

Many courses repeat the same assignment shape every week.

One course might always put the real spec on an external site. Another might always require the same reading annotation format. Another might use the same problem-source and PDF delivery workflow.

If an agent has to rediscover that every week, you are still doing the orchestration. Canvas Pilot tries to make the repeated pattern durable.

The workflow

scan Canvas -> approval plan -> student approval -> approved workflow -> review-ready output -> REPORT.md

The scan step does not start doing assignments. It writes a plan and stops. Execution only happens after the user approves selected items.

Who it is for

This public preview is for AI power users first.

If you already know how to use Codex, Claude Code, or similar local agents, the product can remove a lot of repeated Canvas coordination. It can make recurring coursework feel close to a one-command workflow: scan, approve, execute, review.

If you do not already know how to operate local agent workflows, this version will feel hard. I would rather be honest about that than pretend it is a polished no-code SaaS.