Grok (xAI)

xAI's Grok models over their OpenAI-compatible API. A bare grok-* model name is all mantis needs — it routes to api.x.ai on its own and maps the reasoning knob per model.

Get an API key

1. Sign in at console.x.ai with your X or xAI account
2. Open the API Keys page and click 'Create API key'
3. Name it and copy the xai-… key (shown once)
4. Add credits under Billing

Create a key ↗ · Pricing ↗

Pay-as-you-go; new accounts usually get a small trial credit.

Enable

export XAI_API_KEY=xai-...        # shell profile — survives forever

or in the terminal — validates the key live before saving:

/enable xai

or pick any locked 🔒 Grok (xAI) model in /models and paste the key inline.

Models

Starter menu (once enabled, /models fetches the provider's full live list):

• grok-4 — 256k window, reasoning always on
• grok-4-fast — 2M window
• grok-3
• grok-3-mini — the one that takes reasoning_effort (low / high)

Switch anytime — /model grok jumps to the flagship, /model grok-3 fuzzy-matches; context carries over.

SDK

from mantis_agent import MantisAgentOptions
options = MantisAgentOptions(model="grok-4")      # routes to api.x.ai; key read from $XAI_API_KEY

The explicit form is the same as any other hosted provider:

from mantis_agent import MantisAgentOptions
options = MantisAgentOptions(
    model="grok-4",
    backend="https://api.x.ai/v1",
)

Notes

• GROK_API_KEY works too (alias). The vendor key wins by host, so a stale OPENAI_API_KEY in the same shell is never sent to xAI.
• reasoning_effort is sent only to the models that accept it — grok-3-mini — and the universal thinking config maps onto it. grok-4, grok-3 and grok-code-fast reason at a fixed level and are never sent the field.
• Streamed reasoning_content deltas surface as thinking blocks, collapsed to one line in the terminal (/thinking show expands them).
• Native tool calling on every model; xAI bills cached prompt tokens at a discount, and mantis keeps the prompt prefix stable to make use of it.