coderay 1.2.2

X-ray your codebase — semantic search, code graphs, file skeletons, and MCP server

CodeRay — X-ray your codebase

CodeRay builds a local code index that gives AI agents a smarter way to explore a codebase — reading only what they need, not whole files.

Runs locally. No LLM. No network. No API key.

The problem

AI agents exploring a codebase default to reading whole files – even when one function is all that's needed. Every unnecessary line burns tokens and floods the context window: driving up API costs and noise with every read.

The root cause is simple: agents know the file paths but no finer location. Without knowing where in a file something lives, they have no choice but to read everything.

CodeRay fixes this. Every tool returns file paths with exact line ranges — so agents locate first, then read only the lines that matter.

How it works

CodeRay exposes three primitives, each returning paths + line ranges:

Tool	Question it answers	What agents get
search	Where is the code that does X?	Relevant chunks with file paths and line ranges
skeleton	What's the shape of this file?	Signatures + docstrings only, each tagged with its line range
impact	What breaks if I change this?	Callers, imports, and inheritors — located by line range

The two-phase flow

1. Locate — run search, skeleton, or impact to find what's needed. Every result includes a file path and a symbol-level line range.
2. Read precisely — use those line ranges to load only the relevant snippet. Skip the rest.

This keeps context windows lean and agent reasoning focused. CodeRay is not a replacement for grep — it fills the gap when exact names are unknown or a map is needed before reading.

Token savings (tiktoken, cl100k_base)

File	Lines	Full read	Skeleton	Savings	% reduction
src/coderay/graph/impact.py	249	2,333	693	3.4×	70%
src/coderay/cli/commands.py	584	4,327	1,906	2.3×	56%
src/coderay/pipeline/indexer.py	408	3,065	1,433	2.1×	53%

Query	Search hit tokens	vs full indexer.py read
"how are files re-indexed on change"	479	~6x cheaper

Tools

Semantic search

Agents search by meaning, not by name — useful when the exact function or class is unknown. Results return file paths with line ranges pointing at relevant chunks. Treat them as candidates: confirm with skeleton or a ranged read before acting. Keep the index fresh with coderay watch or coderay build when the tree drifts.

Blast radius

Shows callers, imports, and inheritance for a symbol before it changes. Each result is tied to a file path and line range — combine with skeleton or ranged reads on those locations when bodies are needed.

Skeleton

Returns signatures and docstrings only — no function bodies. Every block is tagged with its path and line range so subsequent reads can be scoped to exactly those lines. A full file read should happen only when the skeleton isn't enough.

Full read

Same file, raw source — for comparison:

First run

coderay init and coderay build.

Status — coderay status: chunks, branch, commit, schema.

MCP

Same three tools over MCP: search, skeleton (paths and line ranges), and impact—so AI agents can narrow context before full-file reads. Point the server at a checkout whose root contains .coderay.toml (CODERAY_REPO_ROOT below). For tool choice versus a plain read, see AGENTS.md.

which coderay-mcp

{
  "mcpServers": {
    "coderay": {
      "command": "/path/to/.venv/bin/coderay-mcp",
      "args": [],
      "env": { "CODERAY_REPO_ROOT": "${workspaceFolder}" }
    }
  }
}

CODERAY_REPO_ROOT must be the directory that contains .coderay.toml. More detail: mcp_server/README.md.

Features

• Languages — Python, JavaScript, and TypeScript — parsing/README.md
• Multi-repo / monorepo — roots, aliases, optional include subtrees — core/README.md
• Hybrid search — vector + BM25 (RRF), optional boosting — retrieval/README.md
• Embeddings — fastembed (CPU) or MLX on Apple Silicon; defaults to MiniLM L6 for speed — configure BGE in .coderay.toml for stronger (heavier) vectors — embedding/README.md
• Watch — incremental re-index; .coderay.toml is the source of truth for what’s indexed

Install

pipx (no venv):

brew install pipx && pipx install coderay   # macOS
# Linux: python3 -m pip install --user pipx && pipx install coderay

In a project:

python -m venv .venv && source .venv/bin/activate
pip install coderay
# Apple Silicon (optional): pip install "coderay[mlx]"

From source: pip install -e ".[all]" — see CONTRIBUTING.md.

Quick start

cd /path/to/your/project
coderay init
coderay watch
coderay search "how does authentication work"
coderay skeleton src/app/main.py
coderay impact some_symbol

CLI

Command	Description
coderay init	Create .coderay.toml and .coderay/
coderay watch [--quiet]	Re-index on file changes
coderay build [--full]	One-off or full rebuild
coderay search "query"	Semantic search (--top-k, --path-prefix, --no-tests)
coderay skeleton FILE	Signatures (--symbol)
coderay impact SYMBOL	Blast radius (--max-depth)
coderay graph	List edges (--from, --to, --kind)
coderay list	Chunks or per-file summary
coderay status	Index metadata
coderay maintain	Compact LanceDB

Configuration

coderay init writes an annotated .coderay.toml: [index], [search], [graph], [embedder], [watcher]. See module READMEs linked from src/README.md.

Contributing

CONTRIBUTING.md

Accuracy and limitations

Semantic search is approximate (model and chunks matter). No warranty — MIT License. Evaluate on your own codebase.