📡 Idea Radar 每天 8:00 / 18:00 推送 · 18:00 生成精选
reddit · SideProject · 49 分 ← 返回列表

I built Codesearch: a structural code intelligence tool for AI agents

I've been working on Codesearch , a local-first code indexing and search tool designed to help AI agents understand unfamiliar codebases. The problem I wanted to solve: investigating a large codebase with an AI agent often means throwing search results at the model, reading files one by one, and hoping it manages to reconstruct how everything fits together. I wanted to give agents something better than grep. What Codesearch does: Parses source code into structural representations

五维评分

付费意愿
32
需求真实度
48
个人可落地
78
竞争空间
24
新鲜度
82

AI 分析

痛点:AI 编码 agent 面对大型陌生代码库时只能依靠 grep 和逐个读文件,缺少结构化的调用关系、引用和依赖图,既耗费大量 token 又难以准确理解系统全貌。

目标用户:重度使用 Claude Code、Cursor、Gemini 等 AI 编码 agent 的开发者与小团队,尤其是维护 10 万行以上多语言代码库的人

证据:investigating a large codebase with an AI agent often means throwing search results at the model, reading files one by one, and hoping it manages to reconstruct how everything fits together. I wanted to give agents something better than grep.

变现:开源核心索引器用于引流,团队版按席位订阅,提供共享索引、云端增量索引、多仓库联合查询和企业私有化部署;也可做托管 MCP 服务按调用量或索引容量收费

MVP:用 tree-sitter 做 AST 解析并存入 SQLite FTS5,只实现函数与类级别的符号表、调用/引用关系;通过 MCP server 暴露 3 个工具(search_symbol、find_references、get_file_context),先支持 Python 和 TypeScript,一周出索引器、一周接 MCP 并做 demo

竞品:Sourcegraph Cody、Cursor 内置 codebase indexing、Aider repo map、Serena MCP code toolkit、Claude Code 内置文件搜索

摘要:作者发布 Codesearch:本地优先的 AST 结构索引工具,结合语义向量与 SQLite 全文检索,通过 MCP server 暴露给编码 agent。实测索引 6538 个文件生成 72000+ 条结构关系,增量重解析 8 个文件耗时 58 秒。帖子属于自发布,未提及任何付费、定价或收入数据,但「让 agent 结构化理解代码库」这一方向正被大量 agent 工具验证。