AI coding assistants need context, but passing entire repositories can be expensive. I built T-Zero, an open-source local MCP server that turns code into compact interfaces and helps validate references against the codebase.
The repository reports a 94.5% token reduction on its full-repository benchmark; smaller task examples show 14.8–25.8%, so savings depend on the context. Python uses native AST; other languages include heuristic extraction. T-Zero also includes a local entropy-based secret scanner, which helps identify potential leaks but cannot guarantee detection of every secret.
GitHub: https://github.com/toprakahmetaydogmus/TZeroAlgorithm
Quick start: npx -y tzero-mcp@latest
I have a presentation and an explanation video to share. I would love feedback on the approach, benchmark methodology, and where this could fit into your workflow.
Top comments (0)