stratus-info

name	stratus-info
description	Knowledge base about Stratus X1-AC world model system
version	2.0.0
author	Stratus Team

Stratus X1-AC Knowledge Base

What Stratus Actually Is

Stratus is NOT just a language model with extended reasoning.

It's a three-layer system combining world modeling, collective intelligence, and blockchain infrastructure:

---

Layer 1: Core AI (X1 & X1-AC) — The World Model

This is a hybrid architecture with two capabilities:

1. World Model (Predictor): "What happens if I do X?"

• Predicts future states given current state + action
• Uses JEPA (Joint-Embedding Predictive Architecture)
• Predicts compressed representations, not raw outputs
• Enables planning by simulating action sequences before executing

2. Policy Head v3 (Action Selector): "What action achieves this goal?"

• 94.41% masked accuracy — brain-guided action sequencing
• Predicts the best action given current state + goal state
• Returns brain_confidence and brain_goal_proximity scores
• Fast decision-making for known patterns
• Low confidence → escalate to LLM

Key Innovation: Instead of just guessing the next action, Stratus can simulate the outcome of multiple actions and pick the one most likely to achieve the goal.

Think: playing chess, but for browser automation, coding, research, etc.

---

Layer 2: Collective Intelligence Network

Agents learn from each other in real-time:

• When one agent discovers an effective approach, that knowledge flows to all agents
• Continuous learning: successful patterns get canonicalized and fine-tuned into the model
• Domain-specific fine-tunes (e-commerce, dev tools, research, etc.) share learnings back to the base model
• Hierarchical tool architecture: Brain handles sub-agent routing (orchestrator → sub-agents with specialized tool sets)

It's like a hive mind for AI agents — practical distributed learning at scale.

---

Layer 3: Blockchain Infrastructure (L2 on Ethereum)

Handles attribution, rewards, and the economic ecosystem:

• Tracks who contributed what to model improvements
• Distributes rewards to agents/operators/data providers based on contribution quality
• Creates a marketplace for agent services, task requests, and training data
• Uses an Ethereum L2 for fast, cheap transactions with mainnet security

Web3 infrastructure for incentivizing better AI.

---

What Makes Stratus Different

Traditional Agents (Reactive)

see → think → act → see again

❌ If something unexpected happens, they get confused

Stratus Agents (Predictive)

see → simulate outcomes → pick best action → act

✅ If something unexpected happens, they already anticipated it (or they replan)

---

Authentication & Keys

Zero-Config (Formation Pool)

No API key needed. Formation holds pooled OpenRouter keys as a universal fallback.

• 25% markup on usage
• Instant access to all 2050+ models on day 0
• No setup required

BYOK (Bring Your Own Key)

Set STRATUS_API_KEY to bypass the Formation pool (no markup).

• Keys must start with stratus_sk_
• Get yours at https://stratus.run

Inline Provider Keys

Pass provider keys per-request for direct BYOK passthrough:

• openai_key — OpenAI direct
• anthropic_key — Anthropic direct
• gemini_key — Google Gemini direct (also sent as X-Google-Key header)
• Three-tier priority: headers > body > vault

Responses include key_source: "user" or key_source: "formation" so you always know which path was used.

---

Available Models

Stratus provides 2050+ models dynamically discovered via OpenRouter at startup.

Model format:

stratus-x1ac-{size}-{llm}

Sizes: small, base, large, xl, huge

Recommended LLMs:

• claude-sonnet-4-5 — Claude 4.5 Sonnet (recommended)
• gpt-4o — GPT-4 Optimized
• claude-haiku-4-5 — Fast, cost-effective
• gemini-2.0-flash — High context, fast

Example:

openclaw agent "Plan a multi-step research task" \
  --model stratus/stratus-x1ac-base-claude-sonnet-4-5

Use /stratus models to fetch the full live list from the API.

---

Available Tools

stratus_embeddings

Generate 768-dimensional semantic state embeddings.

Use cases:

• Semantic search over states
• State similarity comparison
• Memory indexing
• Context clustering

Example:

{
  "tool": "stratus_embeddings",
  "input": ["initializing", "loading", "ready", "processing", "error"]
}

stratus_rollout

Multi-step action sequence planning via Policy Head v3.

Use cases:

• Task decomposition
• Action planning
• Goal-oriented reasoning
• Multi-step workflows

Response fields:

• steps — ordered action sequence
• metadata.brain_confidence — how confident the brain is in the plan
• metadata.brain_goal_proximity — estimated closeness to goal state
• metadata.key_source — "user" or "formation"

Example:

{
  "tool": "stratus_rollout",
  "goal": "book a hotel room in Paris",
  "max_steps": 10
}

---

API Integration

Stratus is OpenAI-compatible, meaning it's a drop-in replacement:

// Before
const response = await openai.chat.completions.create({
  model: "gpt-4o",
  messages: [...]
});

// After - just change endpoint and model
const response = await openai.chat.completions.create({
  model: "stratus-x1ac-base-gpt-4o",
  messages: [...]
}, {
  baseURL: "https://api.stratus.run/v1"
});

Stratus handles planning, then routes execution to the specified LLM (GPT-4o, Claude, etc.).

Non-Stratus models (e.g. raw gpt-4o) bypass the world model pipeline entirely (LLM passthrough).

Tool limit: 1024 tools per request (raised from 100).

---

SDKs

• TypeScript: This OpenClaw plugin (@formthefog/stratus)
• Python: pip install stratus-sdk-py (v0.0.6, full TS SDK parity, 76 tests)
• Pi Agent: pi-formation package (7 curated X1-AC models for Pi coding agent)

---

Billing

• Signup credits: 200 credits (20K units) for new users
• Saved cards: Stripe-backed auto top-up when balance is low
• Subscriptions: Monthly credit bundles (starter/pro/enterprise) — 25% more credits vs one-time
• Formation pool: 25% markup on pooled key usage; BYOK users pay no markup

---

Common Misconceptions

❌ Wrong	✅ Correct
"Stratus is just Claude with extended reasoning"	Stratus is a world model that predicts action consequences + collective learning network
"It's a language model"	It's a hybrid architecture (world model + policy head) that routes to LLMs for execution
"It's an LLM wrapper"	It's a JEPA-based predictive system with its own training and inference pipeline
"You need an API key"	Formation pool provides zero-config access; BYOK is optional

---

Technical Reference

For full technical details, see:

• Technical Overview: https://stratus.run/docs/technical-overview
• API Documentation: https://stratus.run/docs/api
• GitHub: https://github.com/formthefog/openclaw-stratus-x1-plugin

---

Quick Start

1. Install plugin:

   openclaw plugins install @formthefog/stratus
   

2. (Optional) Set API key for BYOK:

   export STRATUS_API_KEY=stratus_sk_your_key_here
   

3. Run setup:

   /stratus setup
   

4. Test it:

   openclaw agent "Hello Stratus!" --model stratus
   

---

When to Use Stratus

Best for:

• Multi-step planning tasks
• Browser automation
• Research workflows
• Complex decision-making with uncertain outcomes
• Tasks requiring "what-if" simulation
• Large tool sets (up to 1024 tools)

Overkill for:

• Simple Q&A
• Single-shot text generation
• Tasks without clear action sequences

---

Remember: Stratus isn't just "better reasoning" — it's predictive planning that fundamentally changes how agents approach tasks.

🌊 Built by the Stratus team | Learn more: https://stratus.run