Skip to main content
You don’t need a new SDK. Add one line to what you already have. 
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["OPENAI_API_KEY"]
)
That’s it. Works with GPT-4o, Claude, Gemini, DeepSeek, Llama, Grok, Mistral, Qwen, and 2,050+ model combinations via OpenRouter.
The base_url is required — Stratus does not auto-detect it. Set STRATUS_API_KEY to your Stratus key (stratus_sk_live_...). No LLM provider key needed to get started — Formation’s pool handles requests automatically. See Authentication for details and BYOK options.

​
Installation

pip install openai
You already have this. Nothing new to install.

​
Making Calls

​
Chat Completions

response = client.chat.completions.create(
    model="stratus-x1ac-base-gpt-4o",
    messages=[
        {"role": "system", "content": "Current state: checkout page, 3 items in cart"},
        {"role": "user",   "content": "Proceed to checkout"}
    ]
)

print(response.choices[0].message.content)
# → "Click the Proceed to Checkout button"

# Stratus-specific planning metadata
print(response.stratus.action_sequence)  # ['click', 'wait', 'verify']
print(response.stratus.overall_confidence)       # 0.94

​
Messages (Anthropic-style)

Works with Anthropic’s client too: 
import anthropic

client = anthropic.Anthropic(
    api_key=os.environ["STRATUS_API_KEY"],
    base_url="https://api.stratus.run"
)

message = client.messages.create(
    model="stratus-x1ac-base-claude-sonnet-4-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Plan the next steps."}]
)

​
Streaming

with client.chat.completions.stream(
    model="stratus-x1ac-base-gpt-4o",
    messages=[{"role": "user", "content": "Plan the deployment steps."}]
) as stream:
    for chunk in stream:
        content = chunk.choices[0].delta.content or ""
        print(content, end="")

​
Models

stratus-x1ac-small-gpt-4o-mini

Lowest cost. Simple classification, routing, and action selection.

stratus-x1ac-base-gpt-4o

Balanced. Most use cases — web navigation, form completion, task planning.

stratus-x1ac-large-gpt-4o

Complex reasoning. Multi-step workflows, cascade prediction, concurrent coordination.

stratus-x1ac-base-claude-sonnet-4-5

Claude backbone. Best for long-context planning and nuanced state reasoning.
See Models for the full list and credit costs per model.

​
The stratus Response Field

Every response includes planning metadata from the X1 world model: 
response = client.chat.completions.create(...)

response.stratus.action_sequence   # list[str] — predicted action chain
response.stratus.overall_confidence        # float    — 0–1 prediction confidence
response.stratus.planning_time_ms  # int      — world model inference time
Use confidence as a gate before executing: 
result = client.chat.completions.create(...)

if result.stratus.overall_confidence < 0.8:
    # Re-describe state with more detail before proceeding
    print("Low confidence — refine state description")

​
Optional: stratus-sdk-py

For vector compression and trajectory simulation, install the Stratus Python SDK: 
pip install stratus-sdk-py

​
Rollout (Pre-Execution Simulation)

Simulate a full action plan before anything executes: 
from stratus_sdk import StratusClient

stratus = StratusClient(api_key=os.environ["STRATUS_API_KEY"])

plan = stratus.rollout(
    goal="Complete the checkout flow",
    initial_state="Cart page, 3 items, coupon field visible",
    max_steps=8
)

if plan.summary.outcome == "success":
    for pred in plan.predictions:
        print(f"Step {pred.step}: {pred.action.action_name}")
Or call the endpoint directly — no extra package needed: 
import requests

plan = requests.post(
    "https://api.stratus.run/v1/rollout",
    headers={
        "Authorization": f"Bearer {os.environ['STRATUS_API_KEY']}",
        "Content-Type": "application/json"
    },
    json={
        "goal": "Complete the checkout flow",
        "initial_state": "Cart page, 3 items, coupon field visible",
        "max_steps": 8
    }
).json()

​
Vector Compression

Compress embedding vectors 10–20× with 99%+ quality: 
from stratus_sdk import compress, decompress, compress_batch, analyze_quality

# Single vector: 6144 bytes → ~600 bytes
compressed = compress(embedding)
restored = decompress(compressed)

# Batch
compressed = compress_batch(embeddings)

# Verify quality before deploying
report = analyze_quality(embeddings, decompress_batch(compressed))
print(report.summary)  # "GOOD (97.2%). Cosine: 99.23%, Recall@10: 95.8%"

​
Next Steps

Authentication

Set up your Stratus key. Optionally add provider keys to remove the Formation pool markup.

Tutorials

Real-world agents — navigation, cascades, concurrency.

API Reference

Full endpoint docs, parameters, and error codes.