deusdata/codebase-memory-mcp v0.12.0 is a local MCP server that indexes source code into a knowledge graph at the level of functions, classes, imports, and HTTP routes, and serves it to coding agents. It runs as a single static binary written only in C, and this release adds CBM Lens, a tool for inspecting the code graph by hand. It provides a 3D architecture map, graph exploration, and a source reader.
Because code relationships are stored in a graph, you query related information instead of opening files one by one. The README reports using roughly 3,400 tokens compared with about 412,000 tokens for the file-reading approach. It supports 162 languages and 17 MCP tools, and offers semantic type resolution via hybrid LSP for Python, TypeScript, Java, PHP, C#, Go, C, C++, Kotlin, Rust, and Perl. The repository description states that it indexes the average repository in milliseconds.
The paper claims that across a comparison of 31 repositories it used 10x fewer tokens and 2.1x fewer tool calls, but its answer quality was lower at 83% for CBM versus 92% for the file-reading approach. The README also cites indexing the28 million lines of the Linux kernel in about 3 minutes and Cypher queries in under 1 millisecond on an M3 Pro. The repository has 3,559 files, but only 6 of them are benchmark scripts and tests. These numbers alone do not establish an accuracy advantage. It can be seen as a code exploration tool that reduces tokens and calls, but it is hard to claim that it replaces file reading in terms of answer quality.