AI writes the code. CodeTrain makes sure you still understand it. A Socratic tutor breaks a real task into small steps on your own repo. You type every line, every step has to run and pass, and what stuck is provable instead of a completion percentage.
Two steps, about a minute, and the first one is going to catch you. Everything here runs for real in your browser. The tutor's replies are pre-recorded, so it knows the handful of ways this lesson usually goes wrong. The real one reads whatever you actually wrote, and it is free for ten lessons a month.
scores holds six results and
names holds the people they belong to, lined up by position, so Ada scored 72
and Bo scored 98.topThree picks out the three best, and the test that
ships with it passes.before: and
after: lines.const names = ["Ada", "Bo", "Cy", "Di", "Eve", "Fay"]
const scores = [ 72, 98, 64, 91, 87, 55 ]
// each person shown next to their score
const pair = (n, i) => n + " " + scores[i]
const board = () => names.map(pair).join(" ")
console.log("before: ", board())
const top = topThree(scores)
console.log("top three:", top)
console.log("after: ", board())
function topThree(scores) {
scores.sort((a, b) => b - a)
return scores.slice(0, 3)
}This lesson needs JavaScript to run. With it on, you get an editor and a Run button right here.
Paste any public repository URL on the free tier and the lesson is built from the code that is actually in it. Four that make for a good first run, and the kind of question each one turns into:
Why does r.headers['content-type'] find a header the server
spelled Content-Type?
It said: This is exactly how
src/requests/structures.py
implements it.
How does current_app know which app it belongs to when nothing
was passed to it?
It said: Flask's real
_cv_app lives in
src/flask/globals.py and works exactly this way.
How does _.get(obj, 'a[0].b.c') reach into a shape that might
not be there, without throwing?
Middleware runs in order until something calls next(). What
happens to a request when one handler forgets to?
Quotes are verbatim from real lessons built on 5 September 2026. Cards without one had nothing worth quoting that day.
Ten lessons a month on the free plan, from any public repo. Pro runs the same thing against your own checkout on 127.0.0.1, where the code never leaves your machine.
Nothing advanced there until your code actually ran.
That is the
difference between a lesson somebody completed and an engineer who can do the work.
CodeTrain is a Socratic AI coding tutor that teaches developers on their own codebase instead of writing code for them. It breaks a real task into small steps, asks questions, reviews each line you write, and runs your tests. Every step must pass before the lesson advances.
Longer reads: how a Socratic coding tutor works, and learning any codebase in 30 minutes, and how this differs from a general AI study mode.
As AI writes more of the code, the understanding behind it erodes. Juniors ship features they can't debug. Onboarding still takes months. "It works" stops meaning "we know why."
The discovery phase of learning, where understanding actually forms, gets skipped when the answer is one tab away.
Source: Digital Education Council, 2026
Generic courses teach someone else's code. Your patterns, your standards, your gotchas go untaught.
Source: Swimm onboarding survey, via HackerNoon (2020)
Seniors burn their week answering the same questions. Skill doesn't compound; it leaks.
Source: SHRM turnover-cost research
No lecture videos, no toy exercises. A Socratic tutor asks, nudges, reviews what you wrote, and runs the tests, but it never hands you the answer.
Paste any public repo URL on the free tier and the lesson is built from the code that's actually in it. Say what you want ("how the API works", "the auth flow") and it aims there. On Pro, repo mode runs against your own checkout and edits land in your real branch, on 127.0.0.1, so your code stays local. Secrets and gitignored files are never read, in either mode.
The tutor breaks the work into tiny steps in a real editor with an instant Run button. You type the line; it reviews what you wrote and explains what changed.
Stuck? It shrinks the step or reveals the next hint, never the solution. Failing code is the best teaching moment, so it leans in instead of bailing out.
The concepts you got stuck on are scheduled to come back at widening intervals. One click turns a due concept into a fresh drill built from what you actually got wrong, not a replay of the lesson. Syncs across the browser, the agent, and the open skill.
New people pull your senior engineers aside with the same questions, and AI-assisted changes land that nobody can fully explain. CodeTrain turns one hard part of your codebase into a guided path. Each person writes the work and runs the checks, and you get a record of what they did without help. Review and merges stay with your people.
Pick what new people struggle with most, like a tricky service or their first safe change. CodeTrain builds the path from your repo and your standards, and every step has to pass before the next one opens.
The team dashboard shows each person's finished lessons, concepts and unaided work. Per-learner budgets keep spend predictable, and you can bring your own model key. Need SSO, audit logs or self-hosting? Talk to us about Enterprise. We are not SOC 2 certified yet.
We take a few design-partner teams at a time. Tell us about the workflow.
Start free. Upgrade when you want your real repos, your own model key, or your whole team. No card to start.
Buying this for yourself rather than a team? The Free and Pro plans are the same product without the team dashboard and shared billing, and the page for individual developers covers them properly.
Or earn your plan instead. Every account gets a referral link, free tier included, and five people on Pro through yours covers your own. How the Conductor Network works.
No, and it is the one rule it will not bend. The tutor's own instructions say the learner writes every line and that it must never write or paste their solution, not even a single line. Stuck learners get a smaller step or the next hint instead. The full rule set is published here.
Opposite jobs. Copilot and Cursor exist to produce code faster, and they are good at it. CodeTrain exists so the person who shipped that code can still explain and change it in six weeks. One is a typing accelerator, the other is a tutor that refuses to type. Most teams should run both. Where each one fits, including when not to use us.
Yes, two ways. The free downloadable agent runs on your machine and reads your local checkout, so nothing leaves it. On Pro, repo mode runs against your own checkout on 127.0.0.1 and edits land in your real branch. On the free managed tier you can point it at any public repo URL. In every mode, secrets and gitignored files are refused before they are read.
Curated courses cover Python, JavaScript, bash and the maths behind LLMs in pure NumPy. Repo mode is not limited to that list: the lesson is built from whatever is in the code you point it at, and steps run in a real sandbox. Browse every course.
No. The agent runs on your machine and serves the UI on localhost. On the managed plan, only short prompt snippets go to the model. On Bring-Your-Own-Key (Team/Enterprise), model calls go straight to your own provider with your key, and nothing but a license check touches our servers.
Those tools turn documents into courses. In software the documents are the least reliable thing you own; the repo is the source of truth, and it moved this morning. A generated course also can't tell you whether anyone can actually do the work, because nothing executes. It grades a written answer against a rubric and reports a completion percentage. CodeTrain builds the lesson from the code itself, makes the engineer write each step, and runs it. If it doesn't pass, the lesson doesn't advance. Different input, different proof. Full comparison, including when not to use us.
You connect your own API key. Anthropic, Google Gemini, Amazon Bedrock, the Gemini Enterprise Agent Platform, OpenRouter, Ollama and any OpenAI-compatible endpoint all work. Your code and prompts go directly to the model on your account. That is best for security and cost control, and we just provide the software, accounts, and dashboards.
Spin up a lesson on your own repo in two minutes. Free to start.