██╗  ██╗██████╗ ███████╗███████╗██╗     ██╗   ██╗██╗  ██╗
  ██║ ██╔╝██╔══██╗██╔════╝██╔════╝██║     ██║   ██║╚██╗██╔╝
  █████╔╝ ██████╔╝█████╗  █████╗  ██║     ██║   ██║ ╚███╔╝ 
  ██╔═██╗ ██╔══██╗██╔══╝  ██╔══╝  ██║     ██║   ██║ ██╔██╗ 
  ██║  ██╗██║  ██║███████╗██║     ███████╗╚██████╔╝██╔╝ ██╗
  ╚═╝  ╚═╝╚═╝  ╚═╝╚══════╝╚═╝     ╚══════╝ ╚═════╝ ╚═╝  ╚═╝

Modern AI inference is trapped behind proprietary walls: opaque guardrails, stealth system prompts, silent model degradations, and brittle single-provider dependencies. When inference fails or providers throttle, developers and researchers are left stranded with no fallback and no transparency.

Kreflux changes the equation.

We are building an open, self-healing reasoning substrate designed for:

1. Resilient Multi-Provider Inference: Zero single points of failure. Inference dynamically fails over across leading providers (Featherless, Neokens, vLLM, OpenRouter) with session resumption and continuous health telemetry.
2. Full Trace Inspectability: No censored thinking. Chain-of-thought (CoT) reasoning traces are completely visible, inspectable, and streamable down to individual tokens.
3. Zero Hidden Prompts: What you send is what the model sees. Transparent system context with zero proprietary steering or hidden instruction injection.
4. Research-Grade Open Weights: We curate and release open datasets, reasoning evaluations, and benchmarks to the global AI research community.

Kreflux is structured as a coordinated ecosystem of applications, agentic tools, evaluation frameworks, and open datasets.

                      ┌──────────────────────────────────────┐
                      │            Kreflux Hub               │
                      │        (https://kreflux.com)         │
                      └──────────────────┬───────────────────┘
                                         │
                 ┌───────────────────────┴───────────────────────┐
                 │                                               │
                 ▼                                               ▼
   ┌───────────────────────────┐                   ┌───────────────────────────┐
   │        kreflux.com        │                   │        klucid CLI         │
   │    Production AI Chat     │                   │    Coding Agent (npm)     │
   │   Multi-Provider Routing  │                   │    Local-first, BYOK      │
   │   Dynamic Compute Ladder  │                   │    Zero Hidden Prompts    │
   └─────────────┬─────────────┘                   └─────────────┬─────────────┘
                 │                                               │
                 └───────────────────────┬───────────────────────┘
                                         │
                                         ▼
                      ┌──────────────────────────────────────┐
                      │             openkreflux              │
                      │      Inference Engine & Router       │
                      │     Reasoning Benchmark Harness      │
                      └──────────────────┬───────────────────┘
                                         │
                                         ▼
                      ┌──────────────────────────────────────┐
                      │         openkreflux-datasets         │
                      │     Curated CoT Reasoning Traces     │
                      │       Hugging Face: Kreflux-AI       │
                      └──────────────────────────────────────┘

The flagship AI chat and reasoning platform hosted at kreflux.com.

• Dynamic Reasoning Ladder: Dial computational depth from Low to Ultra on demand.
• Multi-Provider Failover: Automatic fallback across Featherless, Neokens, and OpenRouter with resume-on-dropout session persistence.
• Rich Scientific Rendering: Real-time streaming markdown, KaTeX math expressions, and interactive Mermaid architectural diagrams.
• Zero Lock-In: Full exportability of all conversations, reasoning tokens, and tool calls.

A local-first, inspectable CLI coding agent included with Kreflux Plus and Pro (available via npm install -g klucid).

• Zero Hidden System Prompts: Full auditability over injected system instructions and context windows.
• Metered Kreflux Compute: Seamless device auth pairing with the Kreflux backend for shared compute allowances.
• Transparent Execution: Interactive approval loops, local file system sandboxing, and live reasoning inspection directly in your terminal.

The open reasoning engine and evaluation framework.

• Multi-Provider Router: High-throughput routing logic with latency-aware load balancing and synthetic fault-injection testing.
• Evaluation Harness: Automated benchmark runner measuring reasoning drift, hallucination indices, and CoT step efficiency.
• Synthetic Reasoning Pipeline: Reproducible generation pipeline for multi-step algorithmic and logical reasoning.

Curated high-fidelity Chain-of-Thought (CoT) reasoning traces hosted on Hugging Face.

• Provenance-Tracked: Every reasoning trace records generation parameters, base model checkpoints, and step verification.
• Open-Weight Friendly: Optimized for fine-tuning open-weight foundation models (Qwen, Llama, DeepSeek) into specialized reasoning architectures.

Repository	Description	Stack	Status
openkreflux	Open reasoning evaluation framework and routing engine	Python 3.10+, PyTorch, vLLM
openkreflux-datasets	Curated CoT reasoning traces and synthetic instruction sets	Hugging Face Datasets, Parquet, JSONL
.github	Organization profile, community health standards, issue templates	Markdown, GitHub Actions

# Install the core OpenKreflux engine from PyPI
pip install openkreflux

# Or using uv
uv pip install openkreflux

# Or clone the open-source repository
git clone https://github.com/Kreflux/openkreflux.git
cd openkreflux
uv pip install -e ".[dev]"

# Install Klucid globally
npm install -g klucid

# Authorize your device with your Kreflux account
klucid auth login

# Start an interactive, inspectable reasoning session
klucid chat "Analyze src/lib/inference.ts and optimize concurrency"

from datasets import load_dataset

# Load curated CoT reasoning traces from Hugging Face
dataset = load_dataset("Kreflux-AI/openkreflux-cot-traces", split="train")

# Inspect the first reasoning trace step
sample = dataset[0]
print(f"Prompt: {sample['prompt']}")
print(f"Reasoning Ladder Level: {sample['reasoning_rung']}")
print(f"Full CoT Trace:\n{sample['thinking_trace']}")

• 🌐 Web Platform: kreflux.com
• 🤗 Hugging Face: huggingface.co/Kreflux-AI
• 🐙 GitHub: github.com/Kreflux
• 🔒 Security Disclosures: security@kreflux.com
• 📜 Code of Conduct: CODE_OF_CONDUCT.md
• 🤝 Contributing Guidelines: CONTRIBUTING.md