Human approval gates for CrewAI

human_input=True on a Task (and the @human_feedback decorator for Flows). Setting human_input=True on a CrewAI Task makes the agent pause before delivering its final answer and wait for a human to type feedback, which is then fed back into the task as added context. The newer @human_feedback decorator (1.8.0+) does the same for Flow steps. On CrewAI's platform, the equivalent is a webhook HITL flow: the crew pauses in a "Pending Human Input" state, fires a webhook with the execution_id, task_id, and task output, and you resume by calling the API with human_feedback and is_approve.

The local human_input=True and @human_feedback primitives block on stdin/console input by default, so they only work in a terminal, not in a web app, API server, or backend crew. The platform webhook HITL solves the pause but hands you everything after it: the docs specify the webhook payload and the resume parameters, but say nothing about where the approval is stored, how it is routed to the right human, who is allowed to approve, or how the decision is audited. Storage, routing-to-a-human, identity, and the audit trail are left entirely to your application to build.

One wrapper is the whole integration surface — the gate is enforced at runtime, before the tool executes.

pip install pliuz

from crewai.tools import tool
from pliuz import gated

# Wrap the underlying tool function. CrewAI has no Pliuz-specific
# adapter — gated() works on any callable.
@tool("wire_transfer")
@gated(policy="wire-transfers", redact=["account"])
def wire_transfer(account: str, amount_eur: float) -> str:
    # This body only runs AFTER a human approves in Slack
    # or the Pliuz web inbox. Reject raises; the run stops.
    return execute_transfer(account, amount_eur)

# Give the gated tool to your agent as usual.
# Every approve / reject / edit is written to the
# tamper-evident, EU-hosted Pliuz audit trail.

With CrewAI alone, approvals go wherever you build them to go: human_input=True blocks on the console, and the platform webhook hands you an execution_id and task output but leaves storage, routing, identity, and audit to your application. With Pliuz, approvals go to a Slack channel or a web inbox where a named human approves, rejects, or edits the call, and the action is blocked at runtime until they do. Every decision is written to a tamper-evident, append-only, Ed25519-anchored audit trail hosted in the EU, which is what turns a console pause into a production approval gate you can prove.