codebase-memory-mcp reimplements four language servers in C

DeusData's codebase-memory-mcp mirrors four language servers, tsserver, pyright, gopls and rust-analyzer, in one C binary with 32.8K GitHub stars already behind it.

AI coding agents normally shell out to real tsserver or gopls processes to answer "who calls this." DeusData skips that: its own benchmark reports 10x fewer tokens and 2.1x fewer tool calls than file-by-file search, at 83% answer quality across 31 repos.

DeusData calls the binary a Hybrid LSP. It reuses those four servers' type-resolution logic to refine CALLS and USAGE edges in the codebase's knowledge graph. The summary also credits typescript-go, Roslyn and Eclipse JDT among the servers it mirrors.

Vendored tree-sitter grammars cover 158 languages for everything the Hybrid LSP doesn't touch. The whole thing indexes the Linux kernel in 3 minutes, per the project's benchmarks.

Those numbers are self-reported across 31 repos. The repo doesn't say what the other 17% of answers get wrong, or how quality was scored.

The bet on approximate over exact

Roslyn and rust-analyzer exist because getting type resolution right takes years of dedicated work per language. Reimplementing four of them by hand in C bets that agents don't need exact resolution, only fast, cheap answers at a 10x token discount.

If 83% holds up outside DeusData's benchmark set, that's a case for baking approximation into coding tools instead of shelling out to real LSP processes. If it doesn't, the wrong 17% is the cost of skipping years of language-server engineering.