Ollama

The Ollama provider talks to a local (or networked) Ollama server. No API key is required; the server itself is the auth boundary.

Construct

import tako

provider = tako.providers.Ollama(
    model="llama3.2",
    base_url="http://127.0.0.1:11434",
    timeout_secs=120,
)

agent = tako.SingleAgent(provider=provider, max_steps=4)
result = await agent.run("Why are octopuses good problem-solvers?")
print(result.text)

Streaming

Ollama streams as NDJSON over /api/chat. The tako adapter exposes it as ordinary OrchEvent::AssistantText:

async for ev in agent.stream("Explain CRDTs"):
    if isinstance(ev, tako.OrchEvent.AssistantText):
        print(ev.text, end="", flush=True)

Vision

Ollama uses a sibling images: [bare-base64] field on each message rather than a content-block array. The tako adapter handles inline images transparently — you still pass them as ContentPart::Image:

from tako import ChatRequest, ContentPart, Message, Role

req = ChatRequest(messages=[
    Message(role=Role.User, content=[
        ContentPart.text("Describe this:"),
        ContentPart.image(mime="image/jpeg", data_b64=jpeg_b64),
    ]),
])

URL-source images via tako pre-fetch

Ollama's wire format doesn't accept URLs, so for ContentPart::ImageUrl content tako has to fetch the URL itself. This is opt-in and default-off because of the SSRF surface — see URL pre-fetch & SSRF for the full mitigation stack.

provider = tako.providers.Ollama(
    base_url="http://127.0.0.1:11434",
    url_prefetch=True,
    url_prefetch_timeout_secs=10,
    url_prefetch_max_bytes=5 * 1024 * 1024,
    url_prefetch_allow_hosts=["*.internal.corp"],
    url_prefetch_allow_cidrs=["10.0.5.0/24"],
)

See also

• Concepts → URL pre-fetch & SSRF
• Concepts → Vision content