CodeLantern

Every engineer who joins a team, switches projects, or picks up an unfamiliar open-source repo hits the same wall: the code is there, but the story is not. READMEs are stale, architecture diagrams live in someone's head, and the only way to learn "what happens when a user upgrades to Pro" is to grep for hours or interrupt a senior teammate. Onboarding routinely takes weeks, and the people who pay for it are new hires, contributors to open source, reviewers, and the seniors who keep re-explaining the same flows. AI summaries do not fix this on their own, because they hallucinate file names and line numbers, so nobody trusts them enough to act on them.

CodeLantern turns any repository into a guided, verifiable story. Point it at a repo (the demo uses shadcn's Taxonomy, or paste a public GitHub URL). It scans the repo once, sends the most important files to Claude in a single structured tool call, and gets back one JSON model, the CodebaseMap: components and their connections, real user journeys, evidence claims, a 5-step tour, and the files to read first. Then our own code verifies every evidence snippet against the actual source and attaches real line numbers. Claude never gets to invent a line number; anything it cannot prove is marked unverified.

From that one model the app renders an interactive architecture graph, a journey tracer that plays a request node by node with the proving code alongside, an evidence panel with syntax-highlighted source and verified lines highlighted, a blast-radius view of what depends on a component, a suggested first task, and a guided tour that walks a newcomer through all of it in five steps. The same JSON also drives a Remotion-generated video you can watch in the app or download as an MP4 to share.

It solves the problem by replacing weeks of reading and interrupting with a 5-minute tour that is trustworthy, because every claim links to the exact lines in the repo. Claude is called once per repo. After that the tour and video are free to view for everyone who joins, with no API key needed.

It scans the repo once, sends the most important files to Claude in a single structured tool call, and gets back one JSON model, the CodebaseMap: components and their connections, real user journeys, evidence claims, a 5-step tour, and the files to read first. Claude never gets to invent a line number; anything it cannot prove is marked unverified. Claude is called once per repo.