Providers¶

tako exposes one trait — LlmProvider — that every model adapter implements. Provider crates depend only on tako-core plus their vendor SDK; they never depend on each other or on tako-runtime. This keeps the dependency graph shallow and makes adding a new provider a mechanical exercise.

Available providers¶

Provider	Crate	Auth	Streaming	Tool calls	Vision
OpenAI	`tako-providers-openai`	API key	✅ SSE	✅	✅ inline + URL
Anthropic	`tako-providers-anthropic`	API key	✅ SSE	✅	✅ inline + URL
Azure OpenAI	`tako-providers-azure-openai`	API key	✅ SSE	✅	✅ inline + URL
AWS Bedrock	`tako-providers-bedrock`	AWS chain	✅ ConverseStream	✅	✅ inline + URL via tako pre-fetch
Vertex AI (Gemini)	`tako-providers-vertex`	OAuth2 token	✅ SSE	✅	✅ inline + `gs://` / `https://` `fileData`
Mistral	`tako-providers-mistral`	API key	✅ SSE	✅	✅ inline + URL
Ollama	`tako-providers-ollama`	local socket	✅ NDJSON	✅	✅ inline + URL via tako pre-fetch
HTTP-generic	`tako-providers-http-generic`	template-driven	⚠️ via `StreamConfig`	⚠️ via template	—
`PythonProvider`	`tako-py`	n/a	✅ async-gen	✅	inline only

Choosing a provider¶

For most users the question is "which API surface do I already have credentials for". The tako trait surface is identical across providers, so you can A/B them with a one-line swap.

provider = tako.providers.OpenAI(model="gpt-5", api_key="...")
# vs
provider = tako.providers.AzureOpenAI(
    endpoint="https://my-resource.openai.azure.com",
    deployment="gpt-4o-prod",
    api_key="...",
)

Conductor orchestration takes a map of providers keyed by worker role, so a single deployment can route code workers to Anthropic and math workers to Vertex without forking the orchestrator.

Custom providers¶

If you need a provider that doesn't ship in the workspace (e.g. an internal LLM gateway), you have two options:

http-generic — point a template at any HTTP endpoint that speaks JSON in/out. Streaming via StreamConfig::OpenAiSse or StreamConfig::NdJson with JSON-pointer-based delta extraction. Python facade: tako.providers.HttpGeneric(...).
PythonProvider — implement chat() (and optionally a streaming async_gen) as Python callables. Useful for prototyping or adapting to non-HTTP transports.

Capabilities¶

Each provider exposes a Capabilities struct — max_context_tokens, supports_streaming, supports_tools, supports_vision, supports_json_mode, plus optional per-million-token cost. The orchestrator consults capabilities before dispatching (e.g. it will fall through to provider.chat(...) for a worker whose supports_streaming is false).

Vision content¶

All seven SDK-backed providers handle outbound vision content:

from tako import ChatRequest, ContentPart, Message, Role

req = ChatRequest(messages=[
    Message(role=Role.User, content=[
        ContentPart.text("Describe this image."),
        ContentPart.image(mime="image/jpeg", data_b64=jpeg_bytes_base64),
        # or, on every provider whose vendor server fetches URLs:
        ContentPart.image_url(url="https://example.com/photo.jpg"),
    ]),
])

For Bedrock + Ollama, URL-source images are fetched server-side by tako (the vendor wire formats don't accept URLs); see URL pre-fetch for the SSRF mitigation stack.