Aurora Research Agent

Autonomous AI research scientist -- multi-model, ARIS-powered scientific research agent.

Aurora handles the full research workflow: literature survey, idea discovery, experiment implementation, cross-model review, and paper writing -- all from the command line.

Features

• Multi-model support -- DeepSeek (default), OpenAI, Anthropic, and any OpenAI-compatible API
• Autonomous agent loop -- Tool-using AI that reads/writes files, runs code, searches the web
• ARIS research workflows -- Idea discovery, auto-review loops, paper writing, full pipeline
• White-label -- Branded as "Aurora", never exposes underlying model/provider identity
• Cross-model review -- Executor and reviewer use different models for adversarial quality control
• Resumable runs -- State persistence enables recovery from crashes or context compaction

Quick Start

Prerequisites

• Python 3.12+
• uv (package manager)
• A DeepSeek API key (get one here)

Install

# Install from PyPI
uv tool install aurora-research
# Creates the `aurora-ai` command

# Or try without installing
uvx aurora-research run "Search arXiv for recent papers on LLM reasoning"

Configure

Set your API key:

# Option 1: Environment variable
export DEEPSEEK_API_KEY="sk-your-key-here"

# Option 2: .env file
echo 'DEEPSEEK_API_KEY=sk-your-key-here' > .env

# Option 3: Interactive config
aurora-ai config

Basic Usage

# One-shot task
aurora-ai run "Explain the Transformer architecture with code examples"

# Interactive chat
aurora-ai chat

# Plan-first execution (read-only research, then execute)
aurora-ai plan "Build a PyTorch training pipeline for CIFAR-10"

# List available models
aurora-ai models

Research Workflows

Idea Discovery

aurora-ai idea "efficient attention mechanisms for long-context LLMs" --num-ideas 8

Pipeline: literature survey -> idea generation (8-12 ideas) -> novelty check -> critical review -> ranked report

Auto Review Loop

aurora-ai review --target ./experiment_results/ --rounds 4

Cross-model adversarial review: reviewer evaluates work -> executor fixes issues -> re-review -> repeat until score >= 7/10

Paper Writing

aurora-ai paper "Efficient Attention via Hierarchical Token Pruning" --assure draft

Pipeline: paper plan -> figures -> LaTeX writing -> compilation -> improvement loop

Full Research Pipeline

aurora-ai pipeline "continual learning without catastrophic forgetting" --write

End-to-end: idea discovery -> experiments -> review -> paper

Model Configuration

Aurora uses model aliases -- user-friendly names that hide the underlying provider:

# Use a specific model
aurora-ai run --model pro "Complex reasoning task"

# Add a custom model
aurora-ai config add turbo --provider openai --model gpt-4o --key sk-xxx

Architecture

CLI (aurora-ai) -> Agent Loop (tool-use engine) -> LLM Backend (multi-model)
                         |
              Tools: read, write, edit, bash, glob, grep, web_search, web_fetch
                         |
              Skills: idea-discovery, auto-review, paper-writing, pipeline

Permission Modes

Development

git clone https://github.com/user/aurora-research.git
cd aurora-research
uv sync
uv run pytest

License

MIT