oooscoos/Benzi

Benzi is AI-native code intelligence infrastructure. Claude Code greps; Cursor embeds; Aider maps signatures; Benzi resolves — and answers in O(1). Every language runs its own tree-sitter grammar into the same query map

Summary

Benzi is a compiler-backed code intelligence layer for AI-assisted engineering. It parses repositories with tree-sitter and resolves imports, symbols, call edges, references, inheritance, and data and control flow into a queryable index, then exposes that map to its coding agent or an MCP-compatible client.

What it is useful for

Benzi is useful when an engineer needs repository-wide context rather than isolated search results. For example, get_callers shows the call sites affected by a function change, trace_path connects two functions and the data carried between them, backflow helps locate where an incorrect value originated, and get_hierarchy reports resolved base types and subclasses.

In its own agent mode, Benzi can read, edit, run, and verify code against this map. Edits are checked against a fresh index and rolled back if they break parsing. Benzi also reports the “blast radius” of a change—callers, holders, and relevant tests—before and after applying it.

The core map supports Python, JavaScript, TypeScript, Java, C#, C++, C, Go, Rust, and Ruby. A separate markup engine handles HTML, CSS, DOM-JavaScript, and frontend code embedded in Python strings.

Teams using an existing agent can serve the same index through MCP. However, MCP provides only the compiler map: runtime tracing, guarded edits, blast-radius checks, persistent memory, rollback, and model escalation belong to Benzi’s own agent. Output quality in MCP mode also depends on the connected agent.

There are documented limitations. Python has the deepest analysis and is the only language with the runtime tracer; incremental reindexing is less optimized for C, C++, Rust, and Ruby. Local models are not yet supported, so the agent currently requires a hosted provider and your own API key.

How engineers can use it

For terminal or CI use, the README documents Python 3.11–3.14 on Windows, macOS, and Linux:

pip install benzi
benzi-login --login [email protected]
benzi-login --model deepseek
benzi-login --nonanthropic-key YOUR_KEY

Then run it from a repository:

benzi-headless . "fix the failing test in parser.py"

For Claude Code, register the MCP server with:

claude mcp add --scope user benzi -- benzi-mcp

A VS Code extension provides the same login configuration and agent workflow. In VS Code, the CLI, and MCP modes, the compiler and index run locally; only code snippets read by the agent are sent to the configured model provider. The browser demo is different: it processes public repositories server-side and limits uploads to 2,000 files of 2 MB each.