What is phenixcode?

PhenixCode is a self-hosted RAG (Retrieval-Augmented Generation) assistant designed for developers who want to query their codebases with AI while keeping code private. It indexes code locally using embeddings from local or cloud LLMs, provides fast vector search via HNSWLib, stores metadata in SQLite, and exposes a REST API with JWT authentication. The project offers prebuilt binaries for Linux, macOS, and Windows with service installation scripts for systemd, NSSM, and LaunchAgent. It supports both local models (via llama-server) and cloud APIs (OpenAI, Mistral, Gemini, XAI, OpenRouter, CometAPI). The UI includes a web dashboard for administration and a chat interface. Configuration is managed through a single JSON file with environment variable overrides.

phenixcode alternatives

All self-hosted alternatives to GitHub Copilot →

Should you self-host phenixcode?

phenixcode scores 36/100 on selfhost.directory (how scoring works) and ranks #100 of 120 in AI Assistant. It has 21 GitHub stars.

It shipped 10 stable releases in the past year. The latest, v2.4.2, came out on 19 October 2026. New versions land roughly every 5 days.

Setup is advanced: plan for several services and some ongoing maintenance. It is MIT-licensed and written mostly in C++.

Its closest alternative on selfhost.directory is openhands (87/100, 92 230 stars). It scores higher overall, so compare the two before you pick. See phenixcode vs openhands side by side.

Features

CLI Interface
Fast vector search
Incremental updates
Local Embeddings
REST API
Service Installation
SQLite Metadata Storage
Web admin dashboard

PhenixCode Assistant

A self-hosted, privacy-first RAG and AI assistance system tailored for codebases. Index your code locally, perform hybrid vector + lexical searches, and get context-aware answers using local models or remote LLM providers.

phenixcode releases & version history

I'm on

v2.4.2Release

10/19/2026 · last week

AI summary Hybrid search and auto-routing for faster responses

Release v2.4.2 — Hybrid Retrieval, Embedded Auto-Router & Decoupled Providers

PhenixCode v2.4.2 brings a major architectural upgrade over the v1.2.x series. This release introduces hybrid search (combining HNSW vector embeddings with SQLite FTS5 BM25 keyword retrieval), a sub-millisecond local task classifier for intelligent model auto-routing without API round-trips, Anthropic API protocol support, a comprehensive retrieval evaluation suite (eval), and persistent chat history.

Breaking Changes & Migration Guide

Configuration Split: `settings.json` and `providers.json`

To simplify multi-project workflows, LLM/Embedding API configurations are now decoupled from per-project settings:

providers.json (Global): Defines all embedding and generation endpoints, API keys, and models (OpenAI, Anthropic, or local llama-server).
settings.json (Per Project): Contains project-specific paths, source inclusion rules, chunking thresholds, and database settings.

What you need to do:
If upgrading from v1.2.x, extract your provider entries (embedding_providers, generation_providers) into a separate providers.json in your working directory, or launch the updated Admin Dashboard to configure them automatically. You can specify a custom provider file using the new -w, --providers-config <path> flag or the PHENIXCODE_PROVIDERS_CONFIG environment variable.

Major Highlights

1. Hybrid Search (Vector + BM25 via Reciprocal Rank Fusion)

SQLite FTS5 Integration: Bundled SQLite now has FTS5 enabled natively.
Reciprocal Rank Fusion (RRF): Queries combine dense semantic search (HNSW) and exact keyword matching (BM25) to catch exact function names, variable symbols, and identifiers that semantic vector models often miss.

2. Zero-Latency Embedded Task Auto-Router

Local C++ Classification Engine: Built-in embedded classifier (assets/baseline_v2.bin) replaces external LLM classification round-trips.
Tiered Routing: Instantly categorizes queries into Simple (Tier 1), Medium (Tier 2), or Complex (Tier 3) within ~70–300 µs on CPU, dynamically dispatching queries to the most cost-effective and capable model.

3. Anthropic Protocol & Cloud LLM Expansion

Native Anthropic API Support: Direct integration with Claude models (e.g., Claude 3.5 / 3.7 Sonnet) via /v1/messages alongside standard OpenAI-compatible endpoints.
Expanded Provider Ecosystem: Tested and optimized for xAI Grok, Mistral, Qwen 2.5 Coder, and local llama-server backends.

4. Retrieval Evaluation CLI (`eval`)

Measure and optimize retrieval quality directly against your codebase:

phenixcode-core eval --init <output.json>: Generate ground-truth benchmark datasets from your indexed code.
phenixcode-core eval --list / --detail <doc_id>: Inspect indexed files and chunk distributions.
phenixcode-core eval <dataset.json> [--top <k>]: Run evaluations benchmarking Recall@k, MRR, and NDCG across dense vs. hybrid retrieval modes.

5. Persistent Chat History & UI Upgrades

SQLite Chat Database: Chat sessions in the desktop client and Webview now persist automatically across restarts.
Markdown Improvements: Full support for Markdown tables and improved one-click code copy button UX.
Reasoning Token Telemetry: Visual indicators for reasoning model token output (e.g., DeepSeek-R1, o3-mini) in the chat status bar.
Central Admin Dashboard: New provider management panels, project directory explorers, and live index statistics.

Bug Fixes & Engine Improvements

Robust Chunking: Fixed edge cases when processing large, opaque, or unsplittable data blobs.
SSE Stream Handling: Resolved corner cases where Server-Sent Events (SSE) were split across network packets.
Token Counter Precision: Fixed off-by-one errors and multi-byte UTF-8 handling in the internal tokenizer.
Graceful Fallbacks: Improved error propagation from upstream inference endpoints back to client webviews.