CrewAI with unoblox (India)

CrewAI coordinates several AI agents, each with its own role, into one workflow. Because every agent typically makes its own model calls, a multi-agent crew can run up meaningful token spend quickly; routing that through unoblox means every agent's calls, regardless of which model backs it, settle in rupees on a single monthly GST invoice.

Why model choice per agent matters more in CrewAI

A crew usually has agents with very different jobs: a researcher that reads a lot of text, a writer that produces the final output, and sometimes a reviewer that checks the writer's work. Giving every agent the same expensive model is the easiest way to build a crew and the easiest way to overspend on it. Since unoblox exposes the full catalog behind one key, you can assign a cheaper model to high-volume roles and a stronger one only to the role whose output quality matters most.

Wiring an agent to unoblox

CrewAI agents accept a model configuration that works the same way as any OpenAI-compatible setup:

from crewai import Agent
from langchain_openai import ChatOpenAI

researcher = Agent(
    role="Researcher",
    goal="Collect and summarise source material accurately",
    backstory="Careful, thorough, cost-conscious",
    llm=ChatOpenAI(
        model="deepseek-ai/deepseek-v4-flash",
        base_url="https://api.unoblox.ai/v1",
        api_key="ub-gw-xxxxxxxxxxxxxxxx",
    ),
)

Repeat the pattern for each agent, changing only the model value to match the role.

Suggested split across a typical crew

Agent role	Model	₹ per 1M tokens (input / output)
Researcher, high volume	DeepSeek V4 Flash	₹9.07 / ₹18.14
Drafting agent	Qwen3 235B-A22B	₹9.07 / ₹55.44
Final writer	GPT-4.1	₹201.6 / ₹806.4
Reviewer, quality gate	Claude Sonnet	₹201.6 / ₹1008

Any model you want to test outside this table, see live ₹ pricing on /models before assigning it to a role that runs on every crew execution.

Watching spend as crews scale

A crew that works well on one example can multiply cost fast once it runs on a real backlog, because every agent-to-agent handoff is its own token-consuming call. Log the model and token count per agent while testing, so you know which role is actually driving cost before you schedule the crew to run unattended on a large queue.

Frequently asked questions

Does CrewAI require a special unoblox integration? No — CrewAI agents typically use a LangChain-style chat model wrapper that already supports a custom base_url and api_key, which is all unoblox needs.

Can different agents in the same crew use completely different model families? Yes — each agent's llm is configured independently, so one agent can run on a DeepSeek model while another runs on GPT or Claude, all under the same unoblox key.

Does data stay in India when agents call GPT or Claude? No — those calls run on the external provider's own infrastructure; unoblox's India benefit for them is rupee billing and a GST invoice, not data residency.

How do we stop one runaway agent loop from spending unexpectedly? Set CrewAI's own iteration and token limits per agent, since unoblox bills every call that goes out; the gateway meters usage, it does not cap a misbehaving loop for you.

Is there a free model for testing a new crew end to end? Yes, Qwen3 1.7B is free (₹0), so you can validate the whole crew's logic before switching individual agents to paid models.

Get started in rupees → https://unoblox.ai/sign-in