An Agent Skill (tested mostly on Claude Code; works on any Agent-Skills-compatible agent) that turns any coding question into a hands-on, one-tiny-step-at-a-time tutorial in a clean, responsive web UI — the style of the popular code-learning sites, but driven by your agent and your own editor.

You write the code; Claude asks, nudges, reviews your submission, runs the tests, and explains what changed. It never dumps the solution on you.

• A tiny Python-stdlib web server (zero dependencies) serves the UI and one JSON state file. The server never executes your code.
• A syntax-highlighted editor (vendored Prism overlay — no runtime CDN, works offline; falls back to a plain textarea) with an instant Run button: Python runs in-browser via Pyodide (WASM), JavaScript in a sandboxed worker — so you self-check against a ✓/✗ test checklist before sending. Bash/shell runs in a throwaway, network-disabled container right from the page when docker/podman is available (else the agent runs it for you).
• You work in the browser; clicking Send to tutor wakes the active Claude Code session to review automatically (Claude idles at ~zero token cost while you code). Running locally first keeps tutor turns rare and cheap — but you can Review my attempt anytime to get a full, generated walkthrough of failing code too. Hints are local and never cost a turn.
• Ask a question anytime (no code needed), Run with ⌘/Ctrl+↵, and End session (two-click, can't misfire) for a celebratory recap + confetti.
• Remembers you: a small local profile (~/.codetrain/profile.json) tracks your level, streak, and recurring gaps — tailoring future sessions and resurfacing things you struggled with (spaced repetition). One small read at the start, two small writes at the end; no token bloat.
• Learning checkpoints (opt-in): while you do normal work, Claude can offer a short, tailored micro-session when your actual code hits a teachable moment — one line, accept/deny, strictly throttled (never nags).
• Two modes:
  • Local code — learn/extend your real codebase with your own local agent. Edits land in place on your current branch, so finished code is just part of your work — nothing to redo. (On main/master it offers a branch first.)
  • Sandbox mode — a throwaway project under /tmp/codetrain-* for practice or any concept not tied to a codebase. Touches none of your files.

• An agent that supports Agent Skills (Claude Code recommended)
• python3 (standard library only — nothing to pip install)
• Optional: docker or podman for in-browser bash (otherwise bash is run by the agent)

Use it (any user):

git clone <this-repo> && cd <repo>/codetrain
./install.sh            # copies into ~/.claude/skills/codetrain

Restart Claude Code.

install.sh also offers to make live sessions prompt-free by adding a small, scoped allow-list to your Claude Code settings (previewed before writing; additive — it never touches your other rules). Run it later anytime with python3 ~/.claude/skills/codetrain/app/install-permissions.py.

Develop it (live edits): symlink instead of copy —

ln -s "$PWD/codetrain" ~/.claude/skills/codetrain

Just ask, in Claude Code:

• "teach me this code"
• "walk me through this function step by step"
• "hold my hand through adding X"
• "give me a practice exercise on recursion"

Claude launches the UI and prints a http://127.0.0.1:PORT link. Open it and go.

• Server binds to loopback only (127.0.0.1). Use SSH forwarding remotely.
• It never executes user code; tests run only when Claude explicitly runs them.
• File writes are path-traversal guarded to the session workspace.
• Progress is stored locally under ~/.codetrain/ (profile + session summaries) — never uploaded, no secrets.
• No external network calls — the UI, fonts (self-hosted, OFL), and syntax highlighter are all vendored; nothing phones home and it works fully offline.
• Prompt-free via a scoped, auditable allow-list (app/install-permissions.py), never a blanket "skip permissions" — the right posture for team/enterprise review.

The server + UI are pure python3 + a browser, and all session updates go through app/ctl.sh and plain files — no Claude-only dependencies. Install by placing the codetrain/ folder in your agent's skills directory (Claude Code: ~/.claude/skills/; others: see your agent's docs), or run ./install.sh.

The one Claude-Code-specific nicety is auto-wake (the agent reacts the moment you click Send). On agents without background auto-wake, the flow is identical except you nudge the agent ("check it") after Send.

codetrain/
  SKILL.md                      # the tutor's instructions (auto-loaded by Claude Code)
  references/session-protocol.md# state schema + the auto-review loop
  app/server.py                 # stdlib server: UI + JSON state API
  app/ctl.sh                    # serve/watch/stop/patch/run control (allowlist-able)
  app/patch.py                  # apply small JSON deltas to session.json (token-cheap)
  app/watch.sh                  # event watcher (wakes Claude on submit/question/end)
  app/checkpoint-hook.sh        # OPTIONAL proactive-checkpoint nudge (off by default)
  app/static/                   # index.html, styles.css, app.js
  app/static/editor.js          # Prism overlay editor (textarea fallback)
  app/static/prism.js           # vendored Prism (highlighting, no CDN)
  app/static/runner.js          # client run dispatcher (timeout, pass/fail)
  app/static/pyodide-worker.js  # Python (WASM) execution worker
  app/static/js-worker.js       # JavaScript sandbox worker
  install.sh