You open a repository you didn't write, or one your coding agent wrote for you, and you can't
tell what does what.
Onramp reads the project and builds a guide you can actually learn from. It asks how much time
you have and what you want to do, then plans a route through the code, stop by stop, from "what
is this" down to the internals. Every technical claim links to the exact lines it comes from,
and a script checks that those lines say what the guide says before you ever see it.
Install
npx skills add SanyamWadhwa07/onramp
The skills CLI finds the agents on your machine and
installs Onramp into the ones you pick: Claude Code, Codex, GitHub Copilot, Cursor, Gemini CLI,
Antigravity, Windsurf, OpenCode, Cline and 60+ more. Name them directly with
-a codex -a cursor.
Claude Code plugin
/plugin marketplace add SanyamWadhwa07/onramp
/plugin install onramp-skill@onramp
Git
git clone https://github.com/SanyamWadhwa07/onramp
cp -r onramp/skills/onramp ~/.claude/skills/onramp # Claude Code
cp -r onramp/skills/onramp ~/.agents/skills/onramp # Codex, Cline, Zed and others
Download the skill folder as a zip from the
latest release and unzip it into your
agent's skills folder.
Onramp needs Python 3.8 or newer. There is nothing else to install.
Use
In Claude Code:
/onramp
In any other agent, ask in plain words: "use onramp to explain this repo".
| Ask for | What you get |
|---|
/onramp | The full guide for this repository, opened in your browser |
/onramp quick | Only the 1-hour route: faster and cheaper |
/onramp explain src/auth/session.ts | A deep lesson on one file, folder or function |
/onramp trace "user resets their password" | One action followed through the code, step by step |
/onramp update | The guide refreshed after the code changed |
/onramp tutor | Your agent teaches you the route in chat, with checkpoint questions |
Everything is written to .onramp/ in your repository. Your code is never modified.
| File | What it is |
|---|
index.html | The interactive guide. One file, opens from disk, no server. |
GUIDE.md | The same guide as markdown, for GitHub and for other agents |
guide.json | The source the page is built from |
What you get
A route, not a pile. Pick your time (1 hour, 1 day, 3 days, 1 week), your goal (get the gist,
ship a change, own it, rebuild it) and your background. Onramp plans a day-by-day route and draws
it as a road with exits, rest stops, checkpoints and a "you are here" marker. Longer routes add
hands-on stops: reading key files in full, doing exercises, building milestones yourself. Your
progress is saved in the browser.
Every claim checked. A citation names a file, a line range and a short piece of text from
those lines, like [[app/store.py:29-38|def place_order(]]. verify.py opens every cited file
and checks the text is really there. Drifted line numbers are re-anchored; a claim whose text is
nowhere in the file fails and is rewritten from the code. Click a citation and the code you see
is copied from your file, never written by the model.
Why it is built this way. Onramp reads the git history, and with your permission your local
Claude Code transcripts, so decisions come with evidence: a commit, a doc, a comment, the prompt
that started it. Each decision is labelled documented or inferred.
Twelve chapters, from the gist to the internals. The gist · Run it · The stack · The map ·
Core ideas · How it flows · Data and state · File by file · How it was built · Why it's like this ·
Handle with care · Your first changes.
Live demos
| Demo | What it is | Guide |
|---|
| PaperMind | A 31,575-line FastAPI + React research-paper app, built with an AI agent | open · 895/895 citations verified |
| nanoGPT | Andrej Karpathy's minimal GPT training repository | open · 496/496 citations verified |
| Flask | The Python web framework, at the commit the evaluation uses | open · 701/701 citations verified |
The PaperMind guide also shows what an honest guide looks like: its "Handle with care" chapter
cites real bugs in the code, such as upload retries that can never succeed and grounding results
that are computed and then dropped.
Evaluation
Onramp is measured with three established methods. Every prompt, answer and grade is in
evals/, and the method is written up in evals/README.md.
| Question | Method | Result |
|---|
| Does the checker catch broken citations? | Fault injection on 2 real guides | 525 / 525 planted faults caught, 0 false alarms on 200 real citations |
| Are the guide's statements about the code true? | FActScore (Min et al., EMNLP 2023), blind, evidence-checked | 96.7% of sampled claims supported (baseline document: 100%) |
| Does the guide help answer real questions? | SWE-QA, 48 Flask questions (ACL Findings 2026), blind | 67.1 / 100, level with the model's own knowledge of Flask (67.2), ahead of the README (63.6) |
Flask is a codebase the models already know well, so the comparison is conservative. The full
write-up, including what the numbers do not show, is in evals/README.md.
How it works
flowchart LR
scan["scan.py<br/>inventory + import graph"] --> read["your agent reads the code<br/>in importance order"]
read --> guide["guide.json<br/>lessons, flows, files"]
guide --> verify["verify.py<br/>checks every citation"]
verify -- "unsupported claim" --> read
verify --> render["render.py"]
render --> out["index.html + GUIDE.md"]
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
