Developer tools · Working system; AWS profile verified locally, not deployed
Unfold
Understand an unfamiliar codebase through evidence, not summaries.
In plain words. Paste a link to someone's code and get a readable guide to how it works, plus a chat where you can ask questions. Every answer points to the exact place in the code it came from, so you can check it yourself.
The problem
A README rarely explains why modules are split, how data moves, or what runs at runtime. A model asked about a repo will happily describe a generic architecture the repo does not have.
What it does
Give Unfold a public GitHub URL or a ZIP and it produces a chaptered guide, architecture views and a chat that answers with citations to files and line ranges in that repository.
Key decisions
- Uploaded code is read as data and never run, which removes a whole class of execution and exfiltration risk.
- Hybrid retrieval: lexical search for exact identifiers, dense vectors for concepts, a symbol graph for relationships.
- Analysis stages are checkpointed, so a provider failure resumes a stage instead of restarting the run and paying for tokens twice.
- Port interfaces keep application logic independent of storage, so a local profile and an AWS profile share one set of contracts.
- Quality regression uses frozen baselines and mutation tests that deliberately corrupt evidence and citations, to prove the gate catches them.
Limits and status
- Static analysis cannot resolve dynamic dispatch, so diagrams are dependency maps, not runtime call graphs.
- Evaluation figures describe specific fixtures and models, not every language or repository.