How RepoLantern turns a repository into a guide
A careful reading of the whole codebase, a cost you approve, and a guide checked against the code before you see it.
AELSHA™ RepoLantern is a free web app and command-line tool that turns any Git repository into a grounded documentation guide: architecture diagrams, citations to exact lines of code, and a check of every name it mentions against the code. You use your own AI model key.
Open any stage for the details.
1. Read
Fetch the code at one commit, safely.
RepoLantern fetches the repository at one exact commit, without running any of its code. It lists every file, finds the languages, entry points, manifests and tests, and removes secrets such as keys and passwords before anything is sent to a model. Large repositories are split into parts that fit the model, and every file is summarised once and reused.
You approve the cost
Nothing is spent until you say yes.
Before any spend, you see how many files will be analysed, how long it will take, and the likely, low and high cost for your model, next to your limit. Nothing is spent until you approve.
2. Understand
From files to folders to the whole system.
File summaries are combined into folder and system views: the architecture, components and flows that every section shares.
3. Write
The guides you chose, from one analysis.
From that one analysis, RepoLantern writes the guides you chose: a beginner's guide, a senior engineer's guide and an architect's guide. Each has its own sections, such as a quick start, key concepts, architecture and data flow, module reference, setup and tests, and a glossary. Diagrams are drawn from the code's real imports.
4. Check
Checked against the code before you see it.
Every file path, function and class a guide mentions is looked up in the real code. A section with names that can't be found is rewritten once, and kept only if it improves. Anything still unverified is listed for you. Citations are resolved to exact lines and link to the analysed commit. How the checks work.
Pause and resume
Nothing finished is redone or billed twice.
Progress is saved step by step. If the server restarts, a model limit is reached or you pause, the job resumes where it stopped, on the same model or another one, without redoing or paying again for finished work. Limits and pauses.
Read, ask, export
Read the guide at three depths, ask questions beside it, and export it to Markdown, PDF, Word, PowerPoint, HTML or draw.io.
Engineered with AELSHA™ ASTRA
RepoLantern is built the way AELSHA™ builds agents for production: as a software system first, with every step observable, measured and gated. You can see this engineering in the app while a guide is made.
- A lifecycle with gates. Each run moves through framing (you approve the scope and cost), building (every part is checkpointed, so a pause or a model swap never redoes or re-bills work), evaluation (names, paths and citations are checked in code and scored), release (only a scored guide is published, to you alone) and observation (every call is traced and costed). Hardening runs throughout: secrets are removed before any model call, text aimed at AI tools is flagged, and spending is capped.
- Measured, not assumed. Every run reports five kinds of signal: telemetry (each model and tool call), evaluation (the grounding score), quality and safety (redactions, flagged text, repairs), cost (spend and cache savings) and outcome (your rating). "How it was made" shows them for each guide.
- Code before judges. Where a check can be code, it is code: twelve frozen golden guides are reproduced exactly, and every planted false name is caught before a release ships. How the checks work.
- Open standards. Traces follow the OpenTelemetry conventions for AI agents, so they can be sent to any tracing tool.
- You stay in control. RepoLantern acts on its own only for low-risk steps; spending, consent and model choice are always yours.