Skip to content

how-it-works.mdx

Latest commit

 

History

History
125 lines (91 loc) · 9.13 KB

how-it-works.mdx

File metadata and controls

125 lines (91 loc) · 9.13 KB
Raw
title How it works
description How RepoLens turns a repo page into a verdict — and the three ways to connect a model, in plain language.

RepoLens is a browser extension that answers one question on any GitHub, GitLab, npm, or PyPI page: should I use this, and what am I signing up for? You click once; it reads the repo, asks an AI model you've connected, and opens a tab that leads with the answer — not the README's sales pitch.

No account. No server. No telemetry. Your keys and your library live in your browser.

A scan, step by step

<svg viewBox="0 0 970 112" role="img" aria-label="Pipeline: click the toolbar icon, read the repo, ask a model, get a verdict, save to your library." style={{ width: '100%', height: 'auto', margin: '10px 0' }}>

<title>How a scan flows, left to right</title> Click the toolbar icon 
<text x="185" y="61" textAnchor="middle" fontSize="20" fill="currentColor" fillOpacity="0.45">→</text>

<rect x="210" y="22" width="150" height="68" rx="12" fill="none" stroke="currentColor" strokeOpacity="0.3" />
<text x="285" y="51" textAnchor="middle" fontSize="14" fontWeight="600" fill="currentColor">Read the repo</text>
<text x="285" y="70" textAnchor="middle" fontSize="11" fill="currentColor" fillOpacity="0.6">README + metadata</text>

<text x="385" y="61" textAnchor="middle" fontSize="20" fill="currentColor" fillOpacity="0.45">→</text>

<rect x="410" y="22" width="150" height="68" rx="12" fill="none" stroke="currentColor" strokeOpacity="0.3" />
<text x="485" y="51" textAnchor="middle" fontSize="14" fontWeight="600" fill="currentColor">Ask a model</text>
<text x="485" y="70" textAnchor="middle" fontSize="11" fill="currentColor" fillOpacity="0.6">the provider you chose</text>

<text x="585" y="61" textAnchor="middle" fontSize="20" fill="currentColor" fillOpacity="0.45">→</text>

<rect x="610" y="22" width="150" height="68" rx="12" fill="#7c5cff" fillOpacity="0.12" stroke="#7c5cff" />
<text x="685" y="51" textAnchor="middle" fontSize="14" fontWeight="600" fill="currentColor">Verdict</text>
<text x="685" y="70" textAnchor="middle" fontSize="11" fill="currentColor" fillOpacity="0.7">fit + bottom line</text>

<text x="785" y="61" textAnchor="middle" fontSize="20" fill="currentColor" fillOpacity="0.45">→</text>

<rect x="810" y="22" width="150" height="68" rx="12" fill="none" stroke="currentColor" strokeOpacity="0.3" />
<text x="885" y="51" textAnchor="middle" fontSize="14" fontWeight="600" fill="currentColor">Library</text>
<text x="885" y="70" textAnchor="middle" fontSize="11" fill="currentColor" fillOpacity="0.6">saved on your machine</text>
  1. Click the RepoLens icon on a repo page.
  2. It reads the public README and metadata (file counts, language, stars, license…).
  3. It asks the model you connected to weigh that evidence.
  4. It opens a verdict — a fit call (strong / solid / care / risky), a one-line bottom line, the flags worth knowing, and where to start.
  5. The result is saved locally to a searchable library you own.

Want the deeper, measured version (real dependency graph, tests, secret scan)? That's the optional runner, used by Deep Dive.

Connecting a model — three ways

RepoLens never ships a key. You bring a model, and there are three shapes that fit different needs:

Way What it means Best when…
Sign-in (OAuth) One click; you approve it on the vendor's site, no key to copy. Available for Claude, Grok, OpenRouter, and OpenAI (Sign in with ChatGPT). The vendor or CLI flow supports OAuth and you'd rather not handle a key.
API key Paste a key from the vendor's console. Works for every provider — Claude, Gemini, OpenAI, DeepSeek, Groq, and ~20 more. You already have a key, or your plan does not expose API access through OAuth.
Local Point at Ollama on localhost. No key, no cloud, no cost. You want it free and private — the AI step never leaves your machine.

Claude has two paths. Use Sign in with Claude for the Claude Code / Pi-style flow, or paste a Console API key (sk-ant-api…). Either way the credential stays in this browser and is excluded from settings exports.

Pick one and you're scanning. See Models for the full provider list, live catalog notes, and per-vendor recommendations.

One key is enough: the smart fallback chain

You don't have to connect everything. Whatever you connect joins a fallback chain — RepoLens tries providers in order and drops to the next if one errors or isn't set up, so a single key (or a local Ollama) is plenty.

<svg viewBox="0 0 970 86" role="img" aria-label="Fallback chain: Nous, then Gemini, then OpenRouter, then Grok, then Anthropic." style={{ width: '100%', height: 'auto', margin: '10px 0' }}>

<title>The default fallback order</title> Nous → Gemini → OpenRouter → Grok → Anthropic

The highest provider in that order that's connected becomes your primary; the rest are automatic backups. Any compatible provider you add (OpenAI, DeepSeek, a local Ollama…) joins as a fallback too.

Want finer control? A scan is really eight parts (core scan, Deep Dive, Versus, …), and you can route each part to a different model — heavy judgment to a big model, bulk re-tagging to a cheap fast one. Each routed part still falls back to the chain, so it can't dead-end.

Run it for $0

Two genuinely-free paths, both first-class:

  • Local Ollama — install Ollama, run a model, and RepoLens does the AI step on your own machine. No key, no bill, fully private.
  • Gemini's free tier — paste a free Google AI Studio key. RepoLens then loads the Gemini models that key can actually call.

(Only the AI step is local with Ollama — RepoLens still reads the public repo page over the network.)

When RepoLens earns its place

  • Evaluating a dependency — "is this library maintained, sane, and worth the lock-in?" before you npm install.
  • Triaging a list — scan ten alternatives; the verdict chips + library let you sort by fit at a glance.
  • Onboarding to a codebase — Deep Dive walks the core concepts in plain English.
  • Comparing two tools — Versus puts them head-to-head on the dimensions that matter.
  • Due diligence — license, health signals, and a secret scan in one place.

Roadmap

Where this is heading. Ideas and directions, not promises — and the order will follow what's actually useful.

Idea What it would do Status
Cost preview Estimate tokens/$ for a scan before you run it. Exploring
Free-starter onboarding First-run flow that points you straight at Ollama or Gemini's free tier. Likely next
Library sync Optional shared/synced store so your library follows you across machines (the storage layer already has the seam for it). Designing
Streaming verdicts Render the verdict as it generates instead of waiting for the whole response. Considering
More one-click sign-ins Add OAuth for any vendor that ships supported third-party OAuth. Ongoing
Team libraries Share a curated, annotated library with a team. Idea

Have a request? RepoLens is meant to be hackable — see Getting started to run it locally.