CrewAI Context Compression with Source Provenance

Install the maintained adapter

pip install 'maha-sdk[crewai]'

The optional extra installs CrewAI support without forcing the framework into the base SDK. The adapter exposes maha_compress_context, maha_verify_claim, and maha_credit_balance. Payment is not autonomous: depleted prepaid credits raise a typed error and require human authorization.

Give the researcher a bounded compression tool

import os
from crewai import Agent, Crew, Task
from maha_sdk import MahaClient
from maha_sdk.crewai import maha_tools

client = MahaClient(api_key=os.environ["MAHA_API_KEY"])
researcher = Agent(
    role="Evidence researcher",
    goal="Answer from source-linked passages within a fixed token budget",
    backstory="Cite retained source passages and state missing evidence.",
    tools=maha_tools(client),
)

task = Task(
    description=(
        "Use maha_compress_context on the supplied documents with a 4000-token "
        "budget. Report the retained evidence and do not invent missing sources."
    ),
    expected_output="A concise answer with source and passage references.",
    agent=researcher,
)

result = Crew(agents=[researcher], tasks=[task]).kickoff(inputs={
    "documents": documents,
})