Eulix

I wrote Eulix to index a repository on your machine, then answer each question from the cheapest path that can: exact lookup, BM25, vectors or the call graph. A model is optional.

Status
◐ Beta
Year
2026
Stack
Rust · Go · Python · MCP · LSP
Source
Nurysso/eulix
37.3M
lines parsed
57 s
Linux kernel, 12 threads
0
parse failures
9
languages

01 / Overview

What it does

Eulix parses your repository into a queryable index and answers from exact lookup, BM25, vector search and call-graph expansion. An LLM is optional and used only where reasoning is needed. Parsed the 37.3M-line Linux kernel in 57 s with 0 failures.

02 / Usage

Three commands

Set up
cd your-project
eulix init      # creates eulix.toml
eulix analyze   # parse and build the index
Ask
eulix chat                      # interactive
eulix query "Who calls foo?"    # one-shot
Give your agent the index
eulix mcp --repo /path/to/repo   # stdio MCP server

03 / Benchmarks

At scale

  • Linux kernel parse

    37.3M lines across 64,460 files in 57 s on 12 threads, with 0 failures. 773K graph nodes and 1.51M edges.

  • Memory

    Spill-to-disk serialization cut peak memory from ~8 GB to ~3.3 GiB.

  • Embeddings

    837,531 chunks embedded locally at ~528 chunks/s on a single consumer GPU (RX 6700 XT).

  • Agent retrieval

    One MCP call returned 30 ranked chunks from 27 files in ~5.4K tokens, replacing an estimated 4–8 grep/read calls.

04 / Pipeline

How it works

  • Index

    libeulix (Rust, Tree-sitter) extracts symbols, imports and call graphs. PRISM, an approximate scope-aware resolver, links the calls. A Python service creates SQ8-quantized embeddings in a memory-mapped vector store.

  • Route

    A deterministic classifier sends each query to the cheapest path: exact lookup, BM25, vector search or call-graph expansion, with MMR re-ranking for diversity.

  • Answer

    Most questions are answered straight from the index. Only open-ended "why" questions go to your LLM, with retrieved code as grounded context.

05 / Ecosystem

What ships

  • eulix (Go CLI)

    Orchestration, query routing, context building and the optional LLM layer.

  • libeulix (Rust)

    The analysis engine as a standalone crate. Supports C, C++, Go, Java, JavaScript, PHP, Python, Rust and TypeScript.

  • eulix mcp (beta)

    Stdio MCP server built into the CLI so coding agents can query your repo.

  • eulix-lsp (beta)

    Language server serving symbol and call-graph analysis to editors.

  • vizEulize

    Browser call-graph explorer with search, filters and neighborhood focus, tested on a 411 MB graph.

06 / Why

Why Eulix

  • Private by default

    Parsing, indexing, retrieval and embedding run on your machine. With a local LLM such as Ollama, nothing leaves it.

  • Any LLM, or none

    Use Ollama, OpenAI, Anthropic, Gemini, DeepSeek or an OpenAI-compatible endpoint. Most queries need no model.

  • Honest scope

    PRISM trades formal soundness for repository-scale speed. Eulix is a navigation and retrieval layer, not a compiler.

07 / Install

Get it running

Linux
bash <(curl -fsSL https://raw.githubusercontent.com/Nurysso/eulix/main/install.sh)
Windows · PowerShell
irm https://raw.githubusercontent.com/Nurysso/eulix/main/install.ps1 | iex

macOS binaries are untested. Open an issue if you hit a problem.

08 / Limits

Rough edges

  • Call resolution is approximate. Dynamic dispatch, reflection and generated code are best-effort.
  • C and C++ parsing is limited by macro expansion, and JSX and TSX are not parsed.
  • The MCP and LSP servers are beta, and macOS binaries are untested.

09 / FAQ

Before you try it

Does my code leave my machine?
Parsing, indexing, retrieval and embedding run on-device. With a local LLM such as Ollama, nothing leaves. If you configure a remote provider, the retrieved code context for each query is sent to it.
Do I need an LLM?
No. Symbol lookups, callers and callees, and project metrics are answered straight from the index. Only open-ended "why" questions go to a model.
Is it a compiler?
No. PRISM is an approximate resolver that trades formal soundness for repository-scale speed. Dynamic dispatch, reflection and generated code are best-effort.
What does it run on?
Linux and Windows. macOS binaries are untested. C and C++ parsing is limited by macro expansion, and JSX and TSX are not parsed.

Eulix

I wrote Eulix to index a repository on your machine, then answer each question from the cheapest path that can: exact lookup, BM25, vectors or the call graph. A model is optional.