Benzi
A compiler-backed AI coding agent that builds a queryable map of calls, data flow, control flow and class hierarchy before reading, editing and verifying code.
At a Glance
Benzi is free to use during beta; proprietary license. CLI and MCP are BYOK, so users pay their own model provider.
Engagement
Available On
Listed Nov 2026
About Benzi
Benzi, from Variant Technologies, is a compiler-first AI coding agent. It compiles a codebase locally into a resolved, queryable index, then uses that index to read, explain, run, edit and verify code. It runs as a VS Code extension or a Python command-line tool, and its compiler can also be served over MCP to other agents. The vendor states it is free to use during beta, with users bringing their own model API key.
What It Is
Benzi is a coding-agent harness built around static analysis rather than grep or embedding search. A tree-sitter-based compiler parses every file, resolves imports, builds class ancestry and traces identifiers to their definitions. The agent then answers questions through structured tools over that index instead of opening files to hunt for context. CodeIndex covers Python, JavaScript, TypeScript, Java, C#, C++, C, Go, Rust and Ruby, and a separate FrontendIndex covers HTML, CSS and DOM-JS.
How the Map and Edit Loop Work
The index exposes three lenses: call flow (who calls whom, including dispatch candidates), data flow (where a value originates or lands) and control flow (under what condition code runs). Every edge carries a confidence label: resolved, candidate, observed, external, unresolved or unindexed.
Edits are gated. Each write is checked by the real language parser and reverts itself if the parse breaks. A landed edit reports the changed symbol, its callers and how the call graph moved. Only changed files are reindexed, and the last N writes can be rolled back from a snapshot. A runtime tracer, documented as Python-only, records real arguments, return values and which dispatch target fired. Per-repo memory tools (remember, recall, forget) keep durable facts across sessions.
Ways to Use It
In VS Code, users install the extension, choose a model, paste an API key (Anthropic, or others such as DeepSeek, Groq and Kimi), log in with an emailed one-time code and run. From a terminal, pip install benzi provides benzi-login, benzi-headless and benzi-mcp for Python 3.11-3.14 on Windows, macOS and Linux. Over MCP, benzi-mcp serves the compiled index to Claude Code, Cursor or any MCP client; the agent-only features (runtime tracer, gated edits, memory, rollback) are not available that way.
Benchmarks Published by the Vendor
Variant reports 413 of 500 (82.6%) on SWE-bench Verified with DeepSeek V4.1-Flash in one attempt per instance, at $9.26 total generation cost. On a separate 24-bug benchmark it reports 9,125 source lines read with Benzi on Sonnet versus 20,704 for Claude Code. These are vendor-published figures.
Privacy and Limits
The compiler and index run on the user's machine and the codebase is not uploaded; only snippets the agent reads go to the chosen model provider. Local models are not yet supported. Language depth is uneven, with Python deepest. The software is proprietary and licensed rather than sold.
Community Discussions
Be the first to start a conversation about Benzi
Share your experience with Benzi, ask questions, or help others learn from your insights.
Pricing
Free (Beta)
Benzi is free to use during beta; proprietary license. CLI and MCP are BYOK, so users pay their own model provider.
- VS Code extension, CLI and MCP compiler
- Bring your own API key (model provider costs paid separately)
- Compiles locally; codebase is not uploaded
- Proprietary license (not open source)
- Free during beta
Capabilities
Key Features
- Compiler-built index of calls, data flow, control flow and class hierarchy
- Ten-language CodeIndex plus HTML/CSS/DOM-JS FrontendIndex
- Confidence-tagged edges (resolved, candidate, observed, external, unresolved, unindexed)
- Parse-gated edits with automatic revert and edit receipts
- Snapshot-based rollback of writes
- Python runtime tracer
- Persistent per-repo memory (remember, recall, forget)
- 36 query tools such as profile, call_tree, trace_path, backflow and skim_source
- Local compilation; codebase not uploaded
- Model-agnostic bring-your-own-key
- Compiler served over MCP
- Self-escalation to a larger model mid-task
