Getting Started

Install SciAgent, learn the current CLI shape, 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.

CLI shape

The current CLI has two top-level verbs:

sciagent run ...
sciagent config ...

Use sciagent run for tasks, and sciagent config show / sciagent config keys to inspect the layered configuration surface.

Installation

Requires Python 3.9+.

pip install sciagent-cli                       # base install (local Docker compute)
pip install 'sciagent-cli[cloud]'              # optional: SkyPilot + AWS extras
pip install 'sciagent-cli[cloud-all]'          # optional: SkyPilot + AWS, GCP, Azure

Or install from source:

python3 -m venv venv
source venv/bin/activate
pip install -e .

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

The agent reads and 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.

--project-dir is optional. If you omit it, SciAgent uses the current working directory.

Interactive mode

For multi-turn conversations:

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

Press Ctrl+C during a task to stop the current run and return to the interactive prompt.

Scientific computing

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

sciagent run --project-dir ~/my-project --task "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 sciagent run options:

Option	Purpose	Default
--task TEXT	Task to execute	—
--project-dir PATH	Directory for reading/writing files	current working directory
--config PATH	Explicit config YAML	—
--set KEY=VAL	Override one config key; repeatable	—
--model NAME	Main model to use (e.g. openai/gpt-4.1)	anthropic/claude-sonnet-4-6
--interactive	Multi-turn conversation mode	Off
--subagents	Enable task orchestration and 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
--system-prompt PATH	Replace the default system prompt	—
--skills-dir PATH	Load SKILL.md definitions from a custom directory	—

Inspect the current surface with:

sciagent --help
sciagent run --help
sciagent config keys
sciagent config show --project-dir ~/my-project