JOIN THE EXPEDITION

Discover what science
hasn't connected yet.

Run MAGELLAN on your machine. The pipeline reads across disciplinary silos, generates testable hypotheses, and publishes the survivors here — attributed to you.

What You Need

Terminal

MAGELLAN runs entirely in the terminal. You need basic familiarity with it — navigating folders, running commands.

Git

Version control, needed to clone the pipeline repository.

Install Git →

Claude Code (CLI)

The terminal-based AI tool that powers MAGELLAN's 12 agents. Requires a Claude subscription (Pro min, Max or Team recommended).

Install Claude Code →

Get Started in 4 Steps

01

Create Your Account

Sign up with your email — no password needed, just a magic link.

Sign In
02

Generate a Contributor Key

Go to your profile and generate a unique contributor key. This links your CLI discoveries to your web profile.

Go to Profile
03

Set Up the Pipeline

Clone the MAGELLAN repo and connect your contributor key. The --enable-auto-mode flag lets the AI agents run autonomously without asking for approval at each step.

git clone https://github.com/kakashi-ventures/magellan-cli.git
cd magellan-cli
claude --enable-auto-mode
/connect mgln_your_key_here
04

Launch a Discovery

Run the discovery pipeline. Results upload automatically when complete.

/discover

Optional: Enable Cross-Model Validation

By default, only Claude validates the hypotheses. For stronger validation, add API keys for GPT-5.4 Pro and Gemini 3.1 Pro so three independent AI models review each hypothesis. Both have free tiers.

1. Create API keys: OpenAI · Google AI Studio

2. Create a .env.local file in the magellan-cli folder:

OPENAI_API_KEY=sk-...
GEMINI_API_KEY=AI...

Without these keys, the pipeline generates export files you can manually copy-paste into ChatGPT or Gemini AI Studio.

Choose Your Path

Scout

Explorer Mode

For the curious. Run /discover with no arguments — the pipeline autonomously picks promising field pairs, generates hypotheses, and attacks them. Fully automatic, 1–2 hours.

/discover
Targeted

Scientist Mode

For domain experts. Specify two fields you want to bridge — the pipeline generates hypotheses at their intersection. Add --context for domain expertise.

/discover Quantum dots × Neural signaling
/discover Soil microbiome × Crop drought --context "I study rhizosphere dynamics"

What Happens When You Run /discover

12

AI agents, 2 cycles

Scout, Generator, Critic, Ranker, Evolver, Quality Gate, and 6 more — each with a specific role, model tier, and output format. The pipeline runs 2 full generation-critique cycles before validation.

30+

web searches per session

Every factual claim is individually verified against PubMed, Semantic Scholar, and KEGG. Citation hallucination or fabricated protein properties = automatic FAIL.

3

AI models must agree

Survivors are independently reviewed by Claude Opus 4.6, GPT-5.4 Pro, and Gemini 3.1 Pro. Only hypotheses that withstand cross-model scrutiny get published.

50+files
~50kwords
500+KB

Each session produces hypothesis cards, critique reports, scoring tables, quality gate audits, cross-model consensus, and meta-learning insights — every claim individually verified against published literature.

FAQ

What is Claude Code?

Claude Code is Anthropic's terminal-based AI coding tool — you interact with it entirely from the command line. MAGELLAN's 12 agents all run through it. It's not the web chat (claude.ai) or the desktop app — it's a CLI you install in your terminal. You need a Claude subscription (Pro minimum, Max or Team recommended for full Opus access). Install instructions at claude.com/product/claude-code.

How much does it cost?

A Claude subscription is the main cost — Pro minimum, Max or Team recommended for Opus access and higher rate limits. With Pro, you may need to reduce model tier and effort levels. Cross-model validation uses OpenAI and Google API keys (both have free tiers) but is entirely optional. OpenAI Codex CLI support is coming soon as an alternative runtime.

How long does a discovery take?

A typical session takes 1–2 hours end-to-end. Cross-model validation (GPT-5.4 Pro + Gemini 3.1 Pro) adds ~30 minutes if API keys are configured. You don't need to watch — type /discover and come back to results.

What exactly does the pipeline produce?

Each session creates 50+ files: hypothesis cards with mechanisms and test protocols, critique reports with kill reasons, 6-dimension scoring tables, quality gate audit with per-claim verification against real papers, cross-model validation consensus from GPT-5.4 and Gemini 3.1, dataset evidence mining, and meta-learning insights. Everything is transparent — you see exactly what was verified, what was killed, and why.

What happens to my results?

Results are saved locally in results/{session-id}/ and automatically uploaded to magellan-discover.ai if you've connected your contributor key. They appear on the public Discoveries page, attributed to your profile. Results are publicly attributed to your profile on the Discoveries page.

Can I run multiple discoveries?

Yes. Each /discover session is independent and explores different scientific territory. The pipeline learns from every session — strategy performance, kill patterns, and bridge survival rates carry over to make future sessions smarter.

What if the upload fails?

Results are always saved locally first. If the upload fails (e.g., you're offline), the admin can sync them manually later. No work is ever lost.