✦ Open source · CLI agent Python ≥ 3.10 · no git, no Node

AI4Science

An installable agent for science — works like Claude Code, then specializes. Chat, let it edit files, run reconstructions, and earn PWM by improving the agents.

View on GitHub ↗ Get started in 5 steps ↓ ⬢ First time? Get PWM → Use → Earn
🤖

Claude Code first

Use it like any coding agent: ask questions, let it edit files (you confirm each change), run one-shot prompts. PWM specialization is opt-in, never the default.

🔬

Science agents & tools

Research, Paper, and Computational-Imaging agents compose verifiable PWM primitives — reconstructions, benchmarks, peer-review panels — via MCP.

⬢

Earn while you use

Using the agent costs PWM; useful /feedback earns it back instantly. Contribute tools/solutions and earn weekly for as long as they're used.

Getting started in 5 steps

Linux/macOS below. Windows: use the PowerShell one-liner on GitHub. Needs only Python ≥ 3.10 — no git, no Node.

1

Install

One line. Open a new terminal afterward so ai4science is on your PATH.

🐧 Linux / 🍎 macOS

curl -fsSL https://raw.githubusercontent.com/integritynoble/AI4Science/main/install.sh | bash

🪟 Windows (PowerShell)

irm https://physicsworldmodel.org/install.ps1 | iex

If a command still isn't found (Linux/macOS): export PATH="$HOME/.local/bin:$PATH"

2

Log in — and you're chatting

Browser approval → PWM founder-served models (no key, no Node). On success it drops you straight into a session.

ai4science login

Prefer your own LLM? ai4science login --provider anthropic|openai|gemini

3

Use it

Start a session any time, or run a one-shot prompt.

ai4science                       # interactive chat session (like `claude`)
ai4science "summarize spec.md"   # one-shot question

Inside a session: ask questions, let it edit files (you confirm each change), /help for slash commands, /exit to leave.

4

Update whenever you like

ai4science update
5

Earn PWM by improving the agents

Using the agent costs PWM; the easiest way to earn it back is feedback. Inside any chat session:

/feedback the CASSI judge should accept .h5 inputs too

Useful feedback is rewarded instantly (an LLM scores its quality). You can also contribute tools/solutions that agents use on paid turns — those earn weekly for as long as they're used. Bootstrap a first balance by mining on physicsworldmodel.org.

Read the full docs & source

Manuals, agent design, the plug-in standard, and earning PWM — all on GitHub.

github.com/integritynoble/AI4Science ↗ 🤖 Browse the agents (MCP)