Providers¶

antcrew (bundled engine) supports 16 LLM providers through a single string prefix. The full list is also available at runtime via GET /engine/supported-providers.

Supported providers¶

Prefix	Provider	BYOK key	Example model string
`claude:`	Anthropic Claude	`anthropic`	`claude:claude-sonnet-5`
`openai:`	OpenAI	`openai`	`gpt-4o`
`gemini:`	Google Gemini	`gemini`	`gemini:gemini-2.0-flash`
`groq:`	Groq	`groq`	`groq:llama-3.1-70b-versatile`
`deepseek:`	DeepSeek	`deepseek`	`deepseek:deepseek-chat`
`mistral:`	Mistral AI	`mistral`	`mistral:mistral-large-latest`
`xai:`	xAI Grok	`xai`	`xai:grok-3`
`together:`	Together AI	`together`	`together:meta-llama/Llama-3-70b-chat-hf`
`fireworks:`	Fireworks AI	`fireworks`	`fireworks:accounts/fireworks/models/llama-v3p1-70b-instruct`
`cerebras:`	Cerebras	`cerebras`	`cerebras:llama3.1-70b`
`moonshot:`	Moonshot AI	`moonshot`	`moonshot:moonshot-v1-8k`
`azure:`	Azure OpenAI	`azure`	`azure:gpt-4o`
`ollama:`	Ollama (local)	—	`ollama:llama3.2`
`lmstudio:`	LM Studio (local)	—	`lmstudio:lmstudio-community/Meta-Llama-3.1-8B-Instruct-GGUF`
`vllm:`	vLLM (local)	—	`vllm:meta-llama/Meta-Llama-3-8B-Instruct`
`simulated`	Simulated (testing)	—	`simulated`

Providers with a BYOK key require an API key stored in your workspace's LLM keys (PATCH /workspaces/{id}/llm-keys). Local providers (Ollama, LM Studio, vLLM) and simulated need no key.

Switching providers¶

from antcrew_engine.config import build_llm

llm = build_llm("claude:claude-sonnet-5")       # Anthropic
llm = build_llm("openai:gpt-4o")               # OpenAI
llm = build_llm("groq:llama-3.1-70b-versatile") # Groq
llm = build_llm("deepseek:deepseek-reasoner")   # DeepSeek R1
llm = build_llm("ollama:llama3.2")             # local Ollama
llm = build_llm("simulated")                   # deterministic mock

BYOK via the platform¶

When using antcrew-platform, store your API keys in workspace settings (Settings → LLM keys) rather than in environment variables. The platform injects the correct key at run time based on the model prefix.

# Store a Groq key for the workspace
curl -X PATCH https://antcrew.org/workspaces/42/llm-keys \
  -H "X-Api-Key: acw_..." \
  -H "Content-Type: application/json" \
  -d '{"groq": "gsk_..."}'

Once stored, any run in the workspace can use groq:* models without setting GROQ_API_KEY in the environment.

BYOK via the proxy¶

When using keybridge, your keys are stored in the proxy and injected at request time. Application code never handles credentials directly:

llm = build_llm("openai:gpt-4o", base_url="https://proxy.antcrew.org")
# proxy injects the real key — your code never handles credentials

Gemini vision calls¶

GeminiModel exposes system_multimodal() for requests that mix text and image data. Useful for code-review capabilities that include screenshots, architecture diagrams, or UI mockups.

from antcrew_engine.models.gemini_model import GeminiModel

llm = GeminiModel("gemini-2.0-flash")
response = llm.system_multimodal(
    system="You are a UI reviewer.",
    parts=[
        {"text": "What accessibility issues do you see in this screenshot?"},
        {"inline_data": {"mime_type": "image/png", "data": "<base64>"}},
    ],
)

system_multimodal() is Gemini-specific. Other providers use the standard complete() method with base64-encoded content in message bodies.

Anthropic extended features (`extra_body`)¶

Pass Anthropic-specific parameters that are not yet in the SDK surface via extra_body:

from antcrew_engine.models.anthropic_model import AnthropicModel

# E.g. inject extended thinking or computer use beta headers
llm = AnthropicModel(
    "claude-sonnet-4-6",
    extra_body={"thinking": {"type": "enabled", "budget_tokens": 5000}},
)

build_llm forwards extra_body to AnthropicModel when the claude: prefix is used:

llm = build_llm("claude:claude-sonnet-4-6", extra_body={"my_key": "value"})

Extra body keys are merged into every API call — both blocking and streaming. Only applies to AnthropicModel.

Discover which providers are active for your workspace¶

The model selector in the Discover page automatically filters to providers your workspace has keys for. From the API:

GET /engine/supported-providers   # full list (16 providers)
GET /workspaces/{id}              # includes byok_providers: ["groq", "deepseek", ...]

byok_providers lists which BYOK keys are configured. Cross-reference with the byok_key field from /engine/supported-providers to know which prefixes are ready to use.