Benzi

Benzi is Variant Technologies’ compiler-backed AI coding agent: before it answers or edits, a tree-sitter compiler parses the repo into one map of symbols, call edges, data flow, control flow, and class hierarchy. The agent then queries that map through ~30 structured tools instead of grepping raw files into the context window.

CodeIndex currently covers Python, JavaScript/TypeScript, Java, C#, C/C++, Go, Rust, and Ruby; FrontendIndex covers HTML, CSS, and DOM-JS. Compilation stays on your machine (code is not uploaded). Run the full agent as a VS Code extension or CLI, or expose just the compiler over MCP to the agent you already use. Free during beta; bring your own model key.

Highlights:

• Every write is syntax-gated by the real parser and auto-reverts on a broken parse;
• A runtime tracer records which call actually fired and with what values;
• Published SWE-bench Verified result: 413/500 (82.6%) on DeepSeek V4.1-Flash, about $9.26 total / 1.85¢ per task;
• Site: https://varianttech.net ; repo: https://github.com/oooscoos/Benzi ; VS Code extension varianttech.benzi.

Why it matters

Most coding agents still guess relevant code via search and embeddings. Benzi treats a compiler map as a first-class interface, claiming less source read per task, blast-radius receipts on edits, and a reproducible SWE-bench number — a complete brand in the “program analysis + agent” line.

Limits

It is still in beta; language coverage and index quality vary with repo complexity. The SWE-bench figure is one model, one pass@1 run, not a guarantee on other models. Compiling a huge monorepo locally costs CPU and disk.

1. Open https://varianttech.net or install Benzi from the VS Code Marketplace (varianttech.benzi); a CLI is also available;
2. Sign in (benzi-login) and add your own model API key;
3. Compile/index the target repo, then let the agent investigate, edit, and run tests;
4. To reuse only the map, run the compiler as an MCP server with your existing agent instead of Benzi’s own loop.