strands-xai

xAI model provider for Strands Agents SDK

Features

• Full Grok Model Support - Access all xAI Grok models (grok-4.5, grok-4.3, grok-4.20-multi-agent, etc.)
• Multi-Agent Research - Orchestrate 4 or 16 collaborating agents with grok-4.20-multi-agent
• Vision Support - Analyze images with vision-capable models (grok-4.5, grok-4.3, grok-build-0.1, grok-4.20-*)
• Server-Side Tools - Use xAI's built-in tools (web_search, x_search, code_execution, collections_search)
• Configurable Reasoning - grok-4.5 supports low/medium/high; grok-4.3 adds none
• Prompt Cache Reporting - Cached input tokens surfaced as cacheReadInputTokens for accurate cost attribution
• Encrypted Reasoning - Preserve sub-agent and reasoning state across turns via use_encrypted_content
• Streaming Support - Real-time response streaming with full event handling
• Hybrid Tool Usage - Combine xAI server-side tools with Strands client-side tools
• Multi-Turn Context - Seamless conversation history with encrypted content preservation
• Type Safe - Full type hints and mypy support

Requirements

• Python 3.10+
• Strands Agents SDK 1.50.2+
• xAI API key from xAI Console

Installation

pip install strands-agents strands-xai

Quick Start

Basic Usage

from strands_xai import xAIModel
from strands import Agent

# Initialize xAI model
model = xAIModel(
    client_args={"api_key": "your-xai-api-key"},  # or set XAI_API_KEY env var
    model_id="grok-4.5",
)

# Create an agent
agent = Agent(model=model)

# Use the agent
result = agent("What is the capital of France?")
print(result)

With Streaming

from strands_xai import xAIModel
from strands import Agent
from strands.handlers.callback_handler import PrintingCallbackHandler

model = xAIModel(
    client_args={"api_key": "your-xai-api-key"},
    model_id="grok-4.5",
)

# Streaming happens automatically with callback handlers
agent = Agent(
    model=model,
    callback_handler=PrintingCallbackHandler()
)

# Text streams to console in real-time
result = agent("Tell me a story")

With Server-Side Tools

from strands_xai import xAIModel
from strands import Agent
from xai_sdk.tools import x_search, web_search

# Use xAI's built-in tools (executed on xAI servers)
model = xAIModel(
    client_args={"api_key": "your-xai-api-key"},
    model_id="grok-4.5",
    xai_tools=[x_search(), web_search()],
)

agent = Agent(model=model)
result = agent("What are people saying about AI on X?")
print(result)

With Reasoning

from strands_xai import xAIModel
from strands import Agent

# Configurable reasoning depth
model = xAIModel(
    client_args={"api_key": "your-xai-api-key"},
    model_id="grok-4.5",
    reasoning_effort="high",  # grok-4.5: "low", "medium", or "high" (default "high")
)

agent = Agent(model=model)
result = agent("Solve this logic puzzle: If all roses are flowers...")
print(result)

Accepted levels differ per model:

Note: grok-4.5 has no none level — it always reasons; pass "low" if you want the cheapest/fastest setting. reasoning_effort is accepted only by grok-4.5 and grok-4.3. The xAI API returns INVALID_ARGUMENT if you pass it to grok-build-0.1, grok-4.20-0309-reasoning, or grok-4.20-0309-non-reasoning — those snapshots have their reasoning behavior baked in. On grok-4.20-multi-agent use agent_count instead — see the Multi-Agent Research section. Reasoning models do not accept presence_penalty, frequency_penalty, or stop in params.

What you get back

For every reasoning-capable model (grok-4.5, grok-4.3, grok-build-0.1, grok-4.20-0309-reasoning, grok-4.20-multi-agent), the SDK streams three signals that this provider surfaces to Strands automatically:

grok-4.20-0309-non-reasoning emits none of the above (reasoning_tokens is zero and the reasoning channels stay empty), as expected for an inference-only snapshot.

With Encrypted Reasoning

For multi-turn conversations with reasoning state preserved across turns:

from strands_xai import xAIModel
from strands import Agent

model = xAIModel(
    client_args={"api_key": "your-xai-api-key"},
    model_id="grok-4.5",
    reasoning_effort="medium",
    use_encrypted_content=True,  # Preserves reasoning across turns
)

agent = Agent(model=model)

# First turn
result1 = agent("Think through this problem: 2+2")
print(result1)

# Second turn - reasoning context preserved
result2 = agent("Now multiply that by 3")
print(result2)

Note: use_encrypted_content is auto-enabled when xai_tools is set so server-side tool state is preserved. You only need to pass it explicitly if you want encrypted reasoning preservation without server-side tools.

Multi-Agent Research (grok-4.20-multi-agent)

Beta: The multi-agent API is currently in beta on xAI. The interface and behavior may change as it iterates — expect occasional breaking changes.

Orchestrate multiple AI agents that collaborate on research tasks:

from strands_xai import xAIModel
from strands import Agent
from xai_sdk.tools import web_search, x_search

# 4 agents for focused queries, 16 for deep research
model = xAIModel(
    client_args={"api_key": "your-xai-api-key"},
    model_id="grok-4.20-multi-agent",
    xai_tools=[web_search(), x_search()],
    agent_count=4,  # or 16 for comprehensive analysis
)

agent = Agent(model=model)
result = agent("Research the latest breakthroughs in quantum computing")
print(result)

Notes:

• The multi-agent model does not support client-side tools (function calling) or max_tokens.
• Only the leader agent's output is returned by default. Sub-agent reasoning, tool calls, and outputs are encrypted and only preserved when use_encrypted_content=True — which xAIModel auto-enables whenever xai_tools is set, so multi-turn context is kept intact.
• All tokens consumed by the leader and every sub-agent are billed (input, output, and reasoning), and each agent's server-side tool calls also count against your tool usage. Expect usage to scale with agent_count.

With Inline Citations

Get sources cited directly in responses:

from strands_xai import xAIModel
from strands import Agent
from xai_sdk.tools import web_search

model = xAIModel(
    client_args={"api_key": "your-xai-api-key"},
    model_id="grok-4.5",
    xai_tools=[web_search()],
    include=["inline_citations"],  # Enable citations
)

agent = Agent(
    model=model,
    system_prompt="You are a helpful assistant. Always cite sources."
)

result = agent("What are the latest developments in AI?")
print(result)
# Output includes inline citations like [1], [2] with source URLs

Vision (Image Understanding)

Analyze images with vision-capable models:

from strands_xai import xAIModel
from strands import Agent

model = xAIModel(
    client_args={"api_key": "your-xai-api-key"},
    model_id="grok-4.5",  # Vision-capable model
)

agent = Agent(model=model)

# Read image bytes
with open("image.png", "rb") as f:
    image_bytes = f.read()

# Send image as content block
message = [
    {"text": "What's in this image?"},
    {"image": {"format": "png", "source": {"bytes": image_bytes}}}
]

result = agent(message)
print(result)

Vision-capable models: grok-4.5, grok-4.3, grok-build-0.1, grok-4.20-0309-reasoning, grok-4.20-0309-non-reasoning, grok-4.20-multi-agent, grok-4.20-multi-agent-0309

Hybrid: Server-Side + Client-Side Tools

from strands_xai import xAIModel
from strands import Agent, tool
from xai_sdk.tools import x_search

@tool
def get_weather(city: str) -> str:
    """Get weather for a city."""
    return f"Weather in {city}: Sunny, 22°C"

# Combine xAI tools with Strands tools
model = xAIModel(
    client_args={"api_key": "your-xai-api-key"},
    model_id="grok-4.5",
    xai_tools=[x_search()],
)

agent = Agent(model=model, tools=[get_weather])
result = agent("What's the weather in Paris and what are people tweeting about it?")
print(result)

Configuration Options

Model Parameters

Common parameters you can pass in params:

model = xAIModel(
    client_args={"api_key": "your-xai-api-key"},
    model_id="grok-4.5",
    params={
        "temperature": 0.7,      # 0.0-2.0, controls randomness
        "max_tokens": 2048,      # Maximum tokens in response
        "top_p": 0.9,            # Nucleus sampling
        "frequency_penalty": 0,  # -2.0 to 2.0
        "presence_penalty": 0,   # -2.0 to 2.0
    }
)

Available Models

xAI's own guidance is to use grok-4.5 for everything text-based, including code. Its knowledge cutoff is February 1, 2026, and it is available in us-east-1 and us-west-2.

Pricing is shown as input/output per million tokens. Dated *-0309 IDs are pinned snapshots; the unsuffixed grok-4.20-multi-agent is a rolling alias that points to the latest stable revision. See xAI documentation for the authoritative pricing and rate-limit reference.

Long-context pricing

Every current Grok model uses tiered long-context pricing: once a request's prompt reaches 200K tokens, all tokens in that request are billed at the higher rate.

Note that grok-4.5 is roughly 1.6× the input cost and 2.4× the output cost of grok-4.3, with a smaller (500K vs 1M) context window — pick it for capability, not for price or raw context length.

Prompt caching

Cached input tokens cost a fraction of fresh input ($0.30 vs $2.00 per MTok on grok-4.5), and this provider reports them to Strands as cacheReadInputTokens so cost backends can apply the discount:

result = agent("...")
print(result.metrics.accumulated_usage)
# {'inputTokens': 4120, 'outputTokens': 380, 'totalTokens': 4500, 'cacheReadInputTokens': 3712}

cacheReadInputTokens is the subset of inputTokens that was served from cache — it is not added on top (the same convention Strands' OpenAI and LiteLLM providers use). It also surfaces on the OpenTelemetry span as gen_ai.usage.cache_read_input_tokens.

xAI recommends pinning a conversation to one server so cache hits are reliable. The gRPC xai_sdk this provider wraps does not expose the Responses-API prompt_cache_key parameter, so there is no strands-xai config option for it; caching still works server-side, but hit rates depend on routing. You can pass the SDK's conversation_id through params for trace grouping:

model = xAIModel(model_id="grok-4.5", params={"conversation_id": "user-42-session-7"})

Keep the stable prefix of your prompt (system prompt, tool definitions) byte-identical across turns — any change to it invalidates the cache.

Model aliases

xAI follows a consistent alias convention for every model line:

So for grok-4.5 the forms are grok-4.5 / grok-4.5-latest; for the multi-agent line they are grok-4.20-multi-agent / grok-4.20-multi-agent-latest / grok-4.20-multi-agent-0309. Two extra aliases resolve to grok-4.5: grok-build-latest (the Grok Build coding agent's default, which is why grok-build-0.1 is effectively superseded) and the generic grok-latest flagship alias. Pass any of these as model_id — the SDK forwards the string verbatim, so you control which side of the rolling/pinned trade-off you take.

Note: xai_sdk's ChatModel type alias does not yet list grok-4.5, but its model parameter is typed ChatModel | str, so passing "grok-4.5" is valid at runtime and under mypy. This provider does not maintain its own allow-list of model IDs — any current or future slug is forwarded as-is.

Retired model aliases

The following slugs were retired on 2026-05-15 and now silently redirect (and are billed) at grok-4.3 or grok-build-0.1 rates. Update your code to migrate explicitly:

Server-Side Tools

xAI provides built-in tools executed on their infrastructure:

Available Tools

• web_search() - Search the web for current information
• x_search() - Search X (Twitter) for posts and trends
• code_execution() - Execute Python code safely
• collections_search() - Search uploaded document collections (RAG)

Basic Usage

from strands_xai import xAIModel
from strands import Agent
from xai_sdk.tools import web_search, x_search, code_execution

model = xAIModel(
    client_args={"api_key": "your-xai-api-key"},
    model_id="grok-4.5",
    xai_tools=[web_search(), x_search(), code_execution()],
)

agent = Agent(model=model)
result = agent("What's trending on X about AI?")

Combining with Client-Side Tools

Mix xAI server-side tools with your own Strands tools:

from strands_xai import xAIModel
from strands import Agent, tool
from xai_sdk.tools import x_search

@tool
def get_weather(city: str) -> str:
    """Get weather for a city."""
    return f"Weather in {city}: Sunny, 22°C"

model = xAIModel(
    client_args={"api_key": "your-xai-api-key"},
    model_id="grok-4.5",
    xai_tools=[x_search()],  # Server-side
)

agent = Agent(
    model=model,
    tools=[get_weather]  # Client-side
)

# Agent can use both types of tools
result = agent("What's the weather in Paris and what are people saying about it on X?")

Observability / OpenTelemetry

Using strands-xai with OpenTelemetry

Strands instruments each model call and emits spans with correct token usage. The underlying xai_sdk also has its own OpenTelemetry instrumentation, which emits a duplicate chat.stream span for the same call. That extra span reports token usage in xAI's raw form (output excludes reasoning, total includes it), which can make trace-level totals inconsistent in observability tools.

When running strands-xai with any OpenTelemetry backend, disable xai_sdk's own tracing so only Strands' spans are exported:

export XAI_SDK_DISABLE_TRACING=1

Set it in your environment before starting the process — xai_sdk binds its tracer at import time, so setting it from Python after import has no effect. With this set, traces contain a single clean span tree and token totals reconcile (input + output == total, with reasoning folded into output).

Cost tracking and the qualified model id

You construct the model exactly as normal — xAIModel(model_id="grok-4.5"). Internally, the provider stores the provider-qualified id xai/grok-4.5 in its config, and that is what appears on the trace as gen_ai.request.model. This is intentional: cost backends (LiteLLM, SideSeat, etc.) price Grok under the qualified key xai/<model> — there is no bare grok-* price key — so the qualified id is what lets them attach a non-zero cost. The xAI SDK itself is always called with the bare id (grok-4.5); the xai/ prefix is stripped at the call site, so inference is unaffected.

You can pass either form ("grok-4.5" or "xai/grok-4.5") — normalization is idempotent. The only observable change is that model.get_config()["model_id"] now returns the qualified xai/grok-4.5.

Examples

See the examples directory for complete working examples.

Interactive Chat

Full-featured interactive chat with 10 different agent configurations:

export XAI_API_KEY="your-xai-api-key"
cd strands-xai
source .venv/bin/activate
python examples/interactive_chat.py

Choose from:

• Simple (non-streaming)
• Streaming with debug mode
• Client-side tools (calculator, weather)
• Server-side tools (X search, web search)
• Hybrid (both server and client tools)
• Reasoning models (grok-4.5 with configurable effort)
• Multi-agent research (grok-4.20-multi-agent)
• Web search with citations

Quick Test

export XAI_API_KEY="your-xai-api-key"
python examples/test_grok_final.py

Or use the convenience script:

./run_examples.sh chat   # Interactive chat
./run_examples.sh test   # Quick test

Development

# Clone the repository
git clone https://github.com/Cerrix/strands-xai.git
cd strands-xai

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Run tests with coverage
pytest --cov=strands_xai --cov-report=html

# Format code
ruff format .

# Lint code
ruff check .

# Type check
mypy src/strands_xai

Testing

Unit Tests

The package includes a comprehensive unit test suite:

# Run all tests
pytest

# Run with coverage
pytest --cov=strands_xai --cov-report=html

# Run specific test
pytest tests/test_xai.py::TestBuildChat -v

Integration Tests with Real API

Test with your xAI API key using the example scripts:

export XAI_API_KEY="your-xai-api-key"

# Interactive testing
python examples/interactive_chat.py

# Quick functionality test
python examples/test_grok_final.py

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Support

• GitHub Issues
• Strands Agents Documentation
• xAI Documentation

Acknowledgments

Built for the Strands Agents community.