Static analysis that compresses code for LLM brains

What it does

TLDR ingests a codebase and builds five layers of static analysis—AST, call graph, control flow, data flow, and program dependence—using tree-sitter. It then queries these layers to generate compact LLM context, trace dependencies, or slice out only the lines relevant to a specific bug. The goal is to stop dumping entire files into a context window and instead hand the model exactly the relevant skeleton.

The interesting bit

Instead of relying on text search or simple file trees, the tool flattens all five analysis layers into 1024-dimensional embeddings using bge-large-en-v1.5, so you can search by behavior—like finding JWT validation logic even if no variable is named jwt. A resident daemon keeps the indexes hot in memory for sub-second queries, and it exposes an MCP server so Claude Desktop can ask questions about your repo directly.

Key highlights

• 95% token reduction for LLM context, according to the project description.
• 16 languages supported, from Python and TypeScript to Elixir and Lua.
• Program slicing isolates exactly the lines that affect a given line number.
• Semantic index built on FAISS lets you search by intent (e.g., “validate JWT”).
• MCP integration for Claude Desktop and Claude Code.

Caveats

• The daemon must be explicitly notified of file changes via CLI calls, git hooks, or editor hooks; it does not appear to watch the filesystem automatically.
• Semantic embeddings are rebuilt only after 20 “dirty” notifications by default, so the index can lag behind active development.

Verdict

Worth a look if you are constantly hitting context-window limits or want an agent to navigate a large legacy repo without reading every line. Less useful if your IDE already manages LLM context to your satisfaction.