Signed decision evidence for CrewAI agents.

Install the native CrewAI integration.

pip install loopgrid-crewai==0.1.0

The package is available on PyPI and integrates with crewai==1.15.25 and the loopgrid 0.8.x SDK. Use an existing LoopGrid Core workspace and the credentials configured for your environment.

Bind one CrewAI Crew run to one LoopGrid decision.

from crewai import Crew, Process
from loopgrid_crewai import LoopGridCrewAI

bridge = LoopGridCrewAI(
    base_url="http://127.0.0.1:8000",
    workspace_id="default",
    agent_id="billing-crew",
)
bridge.install()

decision = bridge.start_decision(
    decision_type="customer_refund",
    agent={"id": "billing-crew", "version": "1"},
    authority={"acting_for": "Example Store", "scope": ["refund:create"], "limit_usd": 100},
    model={"provider": "openai", "name": "configured-model"},
    context={"prompt_version": "refund-v1"},
    proposed_action={"tool": "refund.create", "amount": 25},
    policy={"policy_id": "refund-policy", "version": "1", "decision": "auto_allowed"},
)

# Supply your own real CrewAI agents, tasks and (sandboxed) tools.
crew = Crew(agents=[...], tasks=[...], process=Process.sequential)
with bridge.decision_context(decision["decision_id"]):
    output = crew.kickoff()
bridge.flush()
bridge.assert_healthy()

# Later, only when your application observes a real downstream outcome:
bridge.record_outcome(
    decision["decision_id"],
    {"status": "succeeded", "external_reference": "verified-system-id"},
    observer="billing-webhook",
)

This is an integration template: replace the placeholder agents, tasks and tools with your CrewAI setup. For runnable local scenarios, including native event capture and sandbox tool execution without paid model credentials, use the repository examples ↗.

Observe native CrewAI events; preserve the LoopGrid lifecycle.

Record the action that actually runs.

# Your authenticated approval handler (not the CrewAI tool itself):
bridge.record_human_review(
    decision["decision_id"],
    reviewer="authenticated-reviewer-id",
    approved=True,
    reason="Approved by operator",
)

# Rejection: approved=False, and the action must remain blocked.
# After execution, flush and check delivery:
bridge.flush()
bridge.assert_healthy()

The listener derives paired tool-request and execution records from completed native tool events; it does not claim to intercept a tool before execution. This keeps the evidence consistent with CrewAI's native approval, rejection and argument-modification behavior.

Privacy-safe capture by default.

Use decision_context() to bind the correct decision to a CrewAI run. Native events are queued for delivery without making the CrewAI execution path wait for each network write. Call flush() and assert_healthy() to confirm delivery, and shutdown() when the bridge is no longer needed. Keep credentials, authenticated reviewer identity and observed business outcomes under application control.

Validated from native CrewAI through signed LoopGrid Core.

Validation uses local sandbox tools and a deterministic reviewer fixture, with no real-money movement. LoopGrid verifies recorded evidence integrity/provenance; applications remain responsible for real-world actions and their outcomes.