code-index

Semantic code search for AI coding assistants. Parses your source code, generates LLM-powered summaries, and builds a searchable vector index exposed via MCP.

What it does

code-index lets AI assistants like Claude Code search your codebase by concept instead of exact keywords:

"check if string is in slice"     → finds StringInSlice in util.go
"how does authentication work"    → finds auth packages, token handlers, middleware
"database transaction management" → finds BeginTransaction, CommitTransaction across packages

It works by generating natural language summaries of every function, file, and package, then embedding them as vectors for similarity search.

Supported languages

Quick start

1. Install

Download a pre-built binary from Releases, or install from source:

go install github.com/posit-dev/code-index/cmd/code-index@latest

2. Configure

Create .code-index.json in your repository root (see Configuration):

{
  "project": "my-project",
  "sources": [
    {"path": "src", "language": "go"},
    {"path": "frontend", "language": "typescript"}
  ],
  "llm": {
    "provider": "openai",
    "base_url": "http://localhost:11434/v1",
    "function_model": "llama3.2",
    "summary_model": "llama3.2"
  },
  "embeddings": {
    "provider": "openai",
    "base_url": "http://localhost:11434/v1",
    "model": "nomic-embed-text"
  }
}

This example uses Ollama for fully local operation. You can also use OpenAI, AWS Bedrock, or any OpenAI-compatible API.

3. Build the index

code-index all    # parse → generate → build → embed

4. Add to Claude Code

Add the MCP server to your project's .mcp.json:

{
  "mcpServers": {
    "code-index": {
      "command": "npx",
      "args": ["-y", "@jonyoder/code-index-mcp"]
    }
  }
}

Claude Code will use code_search proactively when working in your codebase.

If your team distributes the vector database via S3 or HTTP URL (see CI Setup), the MCP server automatically downloads it on first search — no manual setup needed.

How it works

parse → generate → build → embed → search
 AST     LLM docs   JSON    vectors   query

1. Parse — extracts functions, types, classes from source files using language-specific parsers
2. Generate — creates LLM summaries for every function (fast model) and file/package (quality model) via AWS Bedrock
3. Build — combines AST data and summaries into a searchable index
4. Embed — generates vector embeddings, stored in a SQLite database with sqlite-vec
5. Search — embeds your query, finds the closest vectors, returns results with signatures, summaries, and source locations

Documentation

• Getting Started — detailed setup walkthrough
• Configuration — full config reference with examples
• AWS Setup — Bedrock access, IAM roles, credential chain
• CI Setup — nightly index updates via GitHub Actions

License

MIT — see LICENSE.