claudix

Copilot's Codebase Index but for Claude Code. Automatically indexes your repo, embeds with the embedding model of choice and provides semantic search through Claude's slash commands, MCP tools, and grep interception.

What It Does

claudix is a Claude Code plugin that gives the agent local semantic search over any repository. A single Rust binary acts as MCP server, hook handler, and CLI. When Claude Code starts, claudix warmly bootstraps the index if missing. When you edit files, chunks are re-embedded automatically. When grep would be less useful than semantic search, the plugin intercepts and uses dense vectors instead. Configuration lives next to settings.json with optional per-project overrides. First-class language support covers Rust, Python, JavaScript, TypeScript, Go, Java, C, C++, C# and SQL.

Core design goal: never break the session, always recover gracefully.

Requirements

• macOS 11+ (Apple Silicon) or Linux x86_64 (glibc 2.28+ or musl) or Windows 10+ x86_64
• Rust 1.91+
• Windows: the Visual C++ 2015-2022 x64 redistributable. The MSVC-built prebuilt binary won't launch without it; a stock box may not have it preinstalled.
• Optional: LM Studio or Ollama for a custom embedding backend. The bundled gte-modernbert-base embedder is statically linked into the binary, so no separate ONNX runtime is needed.

Installation

Requires: Claude Code 2.1.0+ (plugin skills in the slash menu).

claudix ships the bundled gte-modernbert-base embedder and uses it as the fallback when no embedding provider is configured. Set embedding.provider = "http" if you prefer LM Studio, Ollama, or another OpenAI-compatible embedding server.

Linux / macOS

curl -fsSL https://raw.githubusercontent.com/uwuclxdy/claudix/mommy/install.sh | bash

Windows

irm https://raw.githubusercontent.com/uwuclxdy/claudix/mommy/install.bat | iex

Manual

claude plugin marketplace add uwuclxdy/claudix
claude plugin install claudix@claudix

The native binary downloads on first session (~150MB). Restart Claude Code, then run /claudix:doctor to verify.

The downloader honors GITHUB_TOKEN/GH_TOKEN (lifts the anonymous rate limit) and, for mirrors or air-gapped installs, CLAUDIX_RELEASE_API_URL, CLAUDIX_RELEASE_BASE_URL, CLAUDIX_RELEASE_WAIT_MS, CLAUDIX_RELEASE_RETRY_MS.

Targets without a prebuilt (e.g. linux-aarch64, darwin-x86_64) fall back to cargo install claudix@<version>. Install Rust first if you are on one of those.

Configuration

Configuration lives in two TOML files (project overrides global):

• Global: ~/.claude/claudix.toml
• Project: <repo>/.claude/claudix.toml

Both optional. If neither file sets an embedding provider, bundled defaults and the gte-modernbert-base embedder are used. Run /claudix:doctor to see active configuration.

Full Schema (defaults)

watch = false                       # opt-in file watcher; default off (the PostToolUse hook covers edits)

[embedding]
provider = "bundled"                # bundled | http
endpoint = ""                       # required if provider = http (e.g., http://localhost:1234)
model = "gte-modernbert-base"
dimensions = 768
batch_size = 32
timeout_ms = 30000

[indexing]
respect_gitignore = true            # set false to also index gitignored files
follow_symlinks = false
max_file_size_kb = 512              # skip files larger than this
chunk_overlap_lines = 5             # lines each fallback chunk shares with the previous (force-indexed files)
reindex_after_hours = 24            # background reindex interval

[search]
top_k = 10                          # results per search
hybrid_weights = { dense = 0.55, bm25 = 0.30, rrf = 0.15 }  # hybrid retrieval weights
identifier_boost = 1.4              # boost exact identifier matches
similarity_threshold = 0.30         # minimum cosine similarity to keep a candidate
min_score = 0.05                    # minimum fused score to return a hit
cross_repos = []                    # extra already-indexed repo paths to search (read-only)