v1.0 · For Developers & Researchers

Five Intelligences.
One Answer.

5i runs your prompt through five of the world's most powerful AI models simultaneously, then synthesizes their outputs into one neutral, structured consensus — without picking a winner.

GPT-4 · Claude · Gemini · Mistral · Grok

Meet the Five

GPT-4o
OpenAI · Orbit 1
"The overachiever. Gives you three pages when you asked for one sentence. Remarkably good at it though."
Strength: Reasoning, long-form, coding
Style: Thorough, structured, sometimes long-winded
Vibe: Your smartest coworker who emails at 2am
Claude
Anthropic · Orbit 2
"The thoughtful one. Reads the whole paper, considers the ethics, then gives you the nuanced answer you didn't know you needed."
Strength: Nuance, safety, long documents
Style: Careful, balanced, willing to say "it depends"
Vibe: Philosophy PhD who also writes clean code
Gemini
Google · Orbit 3
"Has indexed more of the internet than you've read in your lifetime. Sometimes it shows. Multimodal and unafraid."
Strength: Speed, multimodal, up-to-date knowledge
Style: Confident, snappy, data-forward
Vibe: Google Search grew up and went to grad school
Mistral
Mistral AI · Orbit 4
"The European. Efficient, no-nonsense, surprisingly powerful for its size. Doesn't waste your time or your tokens."
Strength: Efficiency, instruction following, code
Style: Direct, precise, refreshingly concise
Vibe: Berlin engineer who ships twice and ships fast
Grok
xAI · Orbit 5
"The wildcard. Chain-of-thought reasoning out in the open. Might give you the answer, might give you a better question."
Strength: Reasoning, real-time knowledge, contrarian takes
Style: Unfiltered, exploratory, thinks out loud
Vibe: The dev who asks "but why though?" and turns out to be right

How to Run 5i

1

Type your prompt

Enter any question, problem, or research topic in the main input. Works best with open-ended questions where different models might disagree — that's when 5i earns its keep.

// e.g. "What's the best database for a real-time multiplayer game?"
2

Hit Synthesize (⌘↵)

Each model answers sequentially, building on the previous response. You'll see them stream in one by one. The chain runs GPT → Claude → Gemini → Mistral → Grok.

// Watch the orbits glow as each model joins in
3

Read the Consensus

The green block at the top is the Synthesis Engine's output — a neutral aggregation of all five perspectives. The main answer is shown immediately. Hit "Tell me more" to expand the full breakdown: what everyone agreed on, where they diverged, and unique insights from individual models.

// The synthesis never picks a winner. That's the point.
4

Explore the chain outputs

Below the consensus you'll see each model's individual contribution with cost estimates and timing. Useful when you want to see the raw reasoning before it was synthesized.

// "MODEL CHAIN — HOW WE GOT HERE"

The Synthesis Engine

The consensus block isn't a summary. It's produced by a constrained AI instance playing the role of a neutral meta-analyst — not a judge, not a summarizer. It's been explicitly told to behave differently than it normally would.

🔍

Unified Synthesis

A clean, direct answer to the original question — synthesized from all models, not copied from any one of them.

Consensus Points

What all or most models explicitly agreed on. If fewer than 3 models agree, it doesn't count as consensus.

Divergences

Where models meaningfully disagreed. Not smoothed over. Not averaged. Preserved as-is for you to evaluate.

💎

Unique Contributions

Insights that only one model mentioned. Often the most valuable signal — the outlier that turns out to be right.

Open Questions

What none of the models could resolve. Uncertainty is surfaced, not hidden. This is a feature, not a bug.

What the engine is forbidden from doing:
— Picking a "best" answer or ranking models
— Inferring agreement that isn't explicitly stated
— Merging statements that differ in meaning, even if they sound similar
— Discarding minority viewpoints unless they're demonstrably wrong
— Hallucinating consensus where there is none
Why this matters: Standard AI summarizers average outputs or pick a winner. 5i's synthesis engine was designed to be a map of the disagreement space, not a flattener of it. When models agree, that's signal. When they disagree, that's also signal — and you get to see both.

Bring Your Own Keys (BYOK)

5i can run on server-side keys (set by the host) or your own API keys stored locally in your browser. BYOK means your keys never touch the server — they're sent directly with each request and stay in your localStorage.

1

Open Connected Models in the left panel

Click the Set button next to any model. Paste your API key. Done.

2

Badge turns amber → BYOK

Green badge = server key. Amber badge = your key. Your key always takes priority.

3

Keys persist across sessions

Stored in localStorage. They survive page reloads. Use the BYOK button to clear all and start fresh.

Model Provider Get Your Key
GPT-4o OpenAI platform.openai.com/api-keys ↗
Claude Anthropic console.anthropic.com ↗
Gemini Google aistudio.google.com ↗
Mistral Mistral AI console.mistral.ai ↗
Grok xAI console.x.ai ↗
# Health check — see which keys are active on the server
curl https://your-5i-url.railway.app/health

# Expected response when all keys are set:
{"keys": {"anthropic": true, "openai": true, ...}, "status": "ok"}

Keyboard Shortcuts

Synthesize
Toggle System Prompt
/
Randomize Faders
R
Open File
O
Cancel Synthesis
ESC
Copy Output
C

Getting the Most Out of 5i

💡

Ask contentious questions

5i shines when models disagree. "What's the best way to handle auth in a React app?" will get 5 different answers — the divergences section is where the real value lives.

🎚️

Use the faders to weight models

The vertical faders on the left panel adjust each model's weight in the synthesis. Turn down Grok if you want less chaos. Turn up Claude if you want more nuance.

🧠

Use the System Prompt (⌘/) for role-setting

Set a system prompt before synthesizing to frame all models in a specific context — "you are advising a startup CTO" will shift all five responses meaningfully.

📊

Watch the Disagreement Bar

The bar below the Synthesize button shows Agreement, Divergence, Consensus %, and Confidence in real-time. High divergence on a factual question = a model might be wrong.