Codebase Curator

Codebase Curator is an MCP server that turns Claude into a persistent, code‑aware companion for any software project. Instead of treating each question as an isolated prompt, the server creates a dedicated “Curator Claude” that learns the structure, conventions, and hidden dependencies of your entire codebase. This persistent context eliminates repetitive exploration, speeds up responses, and delivers suggestions that are tightly aligned with your team’s coding patterns.

The core problem it solves is the “start‑from‑scratch” cycle that many AI assistants experience when interacting with large repositories. Every query forces the assistant to re‑scan files, rebuild an understanding of architecture, and then answer. This leads to slow, token‑heavy interactions and inconsistent advice that can feel generic or out of sync with your actual code. By spawning a Curator Claude, Codebase Curator establishes an ongoing, deep knowledge cache that remembers previous discoveries and can instantly surface relevant information for follow‑up questions.

Key capabilities are delivered through an intuitive set of tools. The Two‑Claude Architecture creates a separate AI instance that specializes in your project, allowing the main Claude to ask high‑level questions while the Curator handles detailed code analysis. Sessions persist across interactions, so after an initial 2‑minute setup the assistant can answer subsequent queries in a fraction of the time. The Smart Grep tool provides semantic search that understands language constructs across multiple ecosystems (JavaScript, Go, Rust, Swift, shell). It returns concise summaries of definitions, usage sites, and impact analysis with a compact mode that drastically reduces token consumption—making it possible to perform dozens of searches before hitting context limits.

Real‑world use cases include onboarding new developers, performing code reviews, or refactoring large modules. A newcomer can ask, “Where is authentication handled?” and receive a map of all related classes, functions, and configuration values in seconds. A senior engineer can query “What would breaking changes to affect?” and instantly see all downstream references. The Curator’s impact analysis also supports risk‑aware commits by summarizing how uncommitted changes touch the rest of the system.

Integration with existing AI workflows is seamless: developers invoke the Curator via a simple prompt or an API call, and the server returns structured, human‑readable summaries. Because the Curator retains context, developers can chain queries—first discovering a pattern, then asking for best‑practice suggestions—all within the same conversational thread. The result is a fluid, efficient development experience that leverages Claude’s generative power while grounding it in the concrete reality of your codebase.