vexis-crewai 0.1.0

VEXIS AI Governance adapter for CrewAI — task guardrails, step callbacks, and audit trails for every crew execution

vexis-crewai

VEXIS AI Governance adapter for CrewAI — task guardrails, step callbacks, and audit trails for every crew execution.

---

Add one guardrail function. Every task output gets checked against your VEXIS governance policies. PII is caught. Blocked outputs trigger retries. Everything gets an audit trail.

Installation

pip install vexis-crewai

Quick Start: Task Guardrail

from crewai import Agent, Task, Crew
from vexis_crewai import vexis_guardrail

# Create VEXIS guardrail
guardrail = vexis_guardrail(api_key="gp_live_xxx")

# Attach to any task
report_task = Task(
    description="Write a customer analysis report",
    expected_output="A report with customer insights",
    agent=analyst_agent,
    guardrail=guardrail,           # ← VEXIS governance
    guardrail_max_retries=2,       # Retry if blocked
)

If the agent's output contains PII, secrets, or violates your policies, CrewAI automatically retries the task with VEXIS feedback.

Quick Start: Step Callback

from vexis_crewai import VexisStepCallback

callback = VexisStepCallback(api_key="gp_live_xxx")

crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, write_task],
    step_callback=callback,  # ← Every step governed
)

result = crew.kickoff(inputs={"topic": "AI Safety"})

# Governance summary
print(f"Checked: {callback.check_count}, Blocked: {callback.blocked_count}")
print(f"Traces: {callback.trace_ids}")

Combined Governance

from vexis_crewai import VexisCrewGovernance

gov = VexisCrewGovernance(api_key="gp_live_xxx")

task = Task(..., guardrail=gov.task_guardrail())
crew = Crew(..., step_callback=gov.step_callback())

result = crew.kickoff(inputs={...})
print(gov.summary())  # {"checks": 12, "blocked": 1, "trace_ids": [...]}

Configuration

guardrail = vexis_guardrail(
    api_key="gp_live_xxx",
    base_url="https://gateway.internal.corp:8080",  # On-prem
    fail_open=False,          # Block on gateway errors (default)
    metadata={"team": "ml"},  # Extra metadata on traces
)

How It Works

Decision	Guardrail Behavior
ALLOWED	(True, output) — task proceeds
MODIFIED	(True, sanitized_output) — proceeds with PII redacted
BLOCKED	(False, feedback) — CrewAI retries with VEXIS feedback

Links

• Documentation
• VEXIS Dashboard
• GitHub

License

Apache 2.0