codeix

**codeix.dev** · Fast semantic code search for AI agents — find symbols, references, and callers across any codebase.

codeix                 # start MCP server, watch for changes
codeix build           # parse source files, write .codeindex
codeix -r ~/project build  # build from a specific directory

Why

AI coding agents spend most of their token budget *finding* code before they can *work on* it. They grep, read files, grep again, backtrack. On a large codebase the agent might burn thousands of tokens just locating the right function — or worse, miss it entirely and hallucinate.

Codeix gives the agent a pre-built map of your codebase. One structured query returns the symbol name, file, line range, signature, and parent — no scanning, no guessing.

What existing tools get wrong

What codeix does differently

• Committed to git — the index is a `.codeindex` directory you commit with your code. Clone the repo, the index is already there. No re-indexing.
• Shareable — library authors can ship `.codeindex` in their npm/PyPI/crates.io package. Consumers get instant navigation of dependencies.
• Composable — the MCP server auto-discovers dependency indexes and mounts them. Query your code and your dependencies in one place.
• Structured for LLMs — symbols have kinds, signatures, parent relationships, and line ranges. The agent gets exactly what it needs in one tool call instead of piecing it together from raw text.
• Prose search — `search --scope text` targets comments, docstrings, and string literals specifically. Find TODOs, find the error message a user reported, find what a function's docstring says — without noise from code.
• Fast — builds in seconds, queries in milliseconds. Rust + tree-sitter + in-memory SQLite FTS5 under the hood.

The `.codeindex` format

An open, portable format for structured code indexing. Plain JSONL files you commit alongside your code — git-friendly diffs, human-readable with `grep` and `jq`, no binary blobs.

.codeindex/
  index.json        # manifest: version, name, languages
  files.jsonl       # one line per source file (path, lang, hash, line count)
  symbols.jsonl     # one line per symbol (functions, classes, imports, with signatures)
  texts.jsonl       # one line per comment, docstring, string literal

Any tool that can parse JSON can consume a `.codeindex`. Codeix builds it using tree-sitter, and AI agents query it through MCP (Model Context Protocol).

Example — `symbols.jsonl`:

{"file":"src/main.py","name":"os","kind":"import","line":[1,1]}
{"file":"src/main.py","name":"Config","kind":"class","line":[22,45]}
{"file":"src/main.py","name":"Config.__init__","kind":"method","line":[23,30],"parent":"Config","sig":"def __init__(self, path: str, debug: bool = False)"}
{"file":"src/main.py","name":"main","kind":"function","line":[48,60],"sig":"def main(args: list[str]) -> int"}

Ship your index with your package

Include `.codeindex` in your package and every developer who depends on you gets instant navigation of your API — no setup, no re-indexing.

Works with Git repos, npm, PyPI, and crates.io.

MCP tools

Seven tools, zero setup. The agent queries immediately — no init, no config, no refresh.

Project discovery

Launch `codeix` from any directory. It walks downward and treats every directory containing `.git/` as a separate project — each gets its own `.codeindex`.

Works uniformly for single repos, monorepos, sibling repos, and git submodules. No config needed.

Languages

Tree-sitter grammars, feature-gated at compile time:

Markdown support

Markdown files are parsed for headings (both ATX `#` and Setext underline styles) which are indexed as `section` symbols with hierarchical parent-child relationships — enabling TOC extraction and document structure navigation.

Fenced code blocks are extracted as `code` text entries, parented to their containing section.

Embedded scripts

HTML, Vue, Svelte, and Astro files are preprocessed to extract embedded `` blocks, which are then parsed with the JavaScript or TypeScript grammar:

Line numbers in the index point to the original file, not the extracted script block.

Install

# npm / npx — run without installing
npx codeix

# pip / uvx — run without installing
uvx codeix

# Rust
cargo install codeix

# Homebrew
brew install codeix

# Or build from source
git clone https://github.com/montanetech/codeix.git
cd codeix
cargo build --release

All channels install the same single binary. No runtime dependencies.

Usage

# Build the index for the current project
codeix build

# Build from a specific directory (discovers all git repos below)
codeix -r ~/projects build

# Start MCP server (default command, watches for changes)
codeix

# Or explicitly
codeix serve
codeix serve --no-watch

# Serve from a specific directory
codeix -r ~/projects serve

MCP client configuration

Add to your MCP client config (e.g. Claude Desktop, Cursor):

{
  "mcpServers": {
    "codeix": {
      "command": "codeix"
    }
  }
}

Design principles

• Local only — no network, no API keys, works offline and air-gapped
• Deterministic — same source always produces the same index (clean diffs)
• Composable — dependency indexes are auto-discovered and mounted at query time
• Minimal surface — 7 query tools, zero management plumbing

Architecture

See `docs/architecture.md` for the full set of architecture decision records.

License

MIT OR Apache-2.0