This MCP server lets a coding agent query your repo instead of reading it
codebase-memory-mcp jumped onto GitHub trending today, up about 718 stars in a day to 5.1k. It's an MCP server with no LLM of its own, all it does is build a persistent knowledge graph of your codebase and let an agent ask it questions. Tree-sitter AST parsing across 158 languages, hybrid LSP type resolution for the big nine, sub-millisecond queries, and it claims to index an average repo in milliseconds while cutting token usage by 99%.
The design choice is the insight. Today most coding agents understand a repo by reading it, grepping files, stuffing chunks into context, burning tokens to rediscover structure they had a minute ago. This flips it: the agent doesn't read your codebase, it queries a graph of it. Fourteen MCP tools cover search, call-tracing, architecture analysis, and impact detection, the questions an engineer actually asks before touching code.
It ships as a single static binary with zero dependencies on Mac, Linux, and Windows, plugs into eleven agents including Claude Code, Codex, Gemini CLI, and Zed, and even throws in a 3D graph view. Latest is v0.8.1. That's the unglamorous infrastructure layer, not a demo, something you actually wire in.
This is the same drumbeat I keep hearing this month. FastContext split search from solving, turbovec shrank the index, Is Grep All You Need asked whether retrieval even needs vectors. codebase-memory-mcp is the answer that context is the scarce resource and a graph beats grep. As agents get more autonomous the bottleneck stops being the model and becomes how cheaply it can know your code. Repo at https://github.com/DeusData/codebase-memory-mcp.
← Back to all articles
The design choice is the insight. Today most coding agents understand a repo by reading it, grepping files, stuffing chunks into context, burning tokens to rediscover structure they had a minute ago. This flips it: the agent doesn't read your codebase, it queries a graph of it. Fourteen MCP tools cover search, call-tracing, architecture analysis, and impact detection, the questions an engineer actually asks before touching code.
It ships as a single static binary with zero dependencies on Mac, Linux, and Windows, plugs into eleven agents including Claude Code, Codex, Gemini CLI, and Zed, and even throws in a 3D graph view. Latest is v0.8.1. That's the unglamorous infrastructure layer, not a demo, something you actually wire in.
This is the same drumbeat I keep hearing this month. FastContext split search from solving, turbovec shrank the index, Is Grep All You Need asked whether retrieval even needs vectors. codebase-memory-mcp is the answer that context is the scarce resource and a graph beats grep. As agents get more autonomous the bottleneck stops being the model and becomes how cheaply it can know your code. Repo at https://github.com/DeusData/codebase-memory-mcp.
Comments