Getting Started

Install SciAgent and run your first task in minutes.

New in v2.0: cloud compute via SkyPilot, durable provenance log, background subagents with checkpoint/resume. See What’s New in v2.0 for the full list.

Installation

Requires Python 3.9+.

python3 -m venv venv
source venv/bin/activate
pip install -e .                # base install (local Docker compute)
pip install -e '.[cloud]'       # optional: SkyPilot + AWS extras
pip install -e '.[cloud-all]'   # optional: SkyPilot + AWS, GCP, Azure

PyPI package coming soon—for now, install from source.

API Keys

Required - Set your LLM provider key (default model is Claude):

export ANTHROPIC_API_KEY="your-key-here"

Get a key at console.anthropic.com.

Recommended - Add Brave Search for better web results:

export BRAVE_SEARCH_API_KEY="your-key-here"

Get a free key at brave.com/search/api. Without this, web search falls back to DuckDuckGo.

Running your first task

sciagent --project-dir ~/my-project "Create a hello world Python script"

The agent reads/writes files in ~/my-project, runs shell commands, searches the web, and tracks progress with a todo list. When finished, you’ll see a summary and can inspect the generated code.

Interactive mode

For multi-turn conversations:

sciagent --project-dir ~/my-project --interactive

Press Ctrl+C anytime to pause and choose to continue, stop, or redirect.

Scientific computing

For simulations, SciAgent uses containerized services (SciPy, RCWA, MEEP, etc.):

sciagent "Run an RCWA simulation for a photonic crystal grating"

The agent researches documentation, writes code, and runs it in Docker automatically.

Command-line options

Common options (defaults from AgentConfig):

Option	Purpose	Default
--project-dir PATH	Directory for reading/writing files	Required
--model NAME	LLM to use (e.g. openai/gpt-4.1)	anthropic/claude-sonnet-4-6
Fast model	Used for explore subagent and content processing. See defaults.py	anthropic/claude-haiku-4-5-20251001
--interactive	Multi-turn conversation mode	Off
--subagents	Enable WorkflowTool for full DAG execution	Off
--max-iterations N	Max agent loop cycles	120
--temperature T	LLM randomness (0 = deterministic)	0.0
--resume ID	Continue a previous session	—
--list-sessions	Show available resumable sessions	—
--quiet	Minimal output	Off

Run sciagent --help for all options.