Benzi

The codebase, compiled for the agent.

Website  ·  Live demo  ·  Benchmark  ·  GitHub

Benzi is a coding agent that actually understands your codebase — not by grepping or guessing, but by compiling it into a precise, queryable map: every symbol, call edge, and reference resolved up front. It navigates, reads, runs, and edits through that map, so it's fast (it never re-reads your whole repo) and precise (it reasons from proven structure, not vibes). It can even run your code to capture real runtime values, and it remembers what it learns from one session to the next.

Try it in 30 seconds

• Live demo — no install: benzi.fly.dev — paste any public GitHub repo link and start asking.

What makes it different

• A resolved map, not a text search. Benzi compiles your project once into an exact index — proven call edges, data flow, references — and queries that instead of re-reading files. Fewer tokens, sharper answers.
• It runs your code. Click a Python file's run button (or just ask) and Benzi traces the actual execution, capturing real argument and return values — the only truth for dynamic dispatch, callbacks, and anything a runtime decides.
• Everything is verifiable. When Benzi names a function, it lights up on the graph — click to jump straight to the source. Answers you can check, not take on faith.
• Persistent memory. What it learns about your codebase carries across sessions.

Requirements

• Python 3.9+ on your machine — the extension builds its own isolated environment on first run (nothing to configure).
• No API key needed — chat runs on the hosted Benzi service.
• Edits are reversible — every change Benzi makes is confined to your workspace and recorded, so it can be rolled back.

Links

• 🌐 Website — https://benzi.fly.dev/about
• ▶️ Live demo — https://benzi.fly.dev
• 📊 Benchmark — https://benzi.fly.dev/benchmark
• 🐙 GitHub — https://github.com/shobhitx64/Benzi

_{Built by Variant.}