Sqube Agent Control

Open-source execution authorization infrastructure for AI agents.

AI proposes. Sqube decides. Sqube records what actually happened.

v1.0 covers deterministic policy, approvals, tamper-evident ledgers, optional OpenTelemetry, a local control plane (dashboard + APIs), and remote event ingestion (agents export ledger events; the plane does not authorize runtime execution).

Status

v1.0 — Python is the reference SDK; Node and Rust share the synchronous execution contract (tests/contract/). SDK-level control is bypassable if application code skips the guard.

Install

pip install sqube-agent-guard
npm install sqube-agent-guard

Rust: build from the repository rust/ directory (crates.io pending).

Quick start (Python)

from sqube_agent_guard import Decision, ExecutionGuard

def policy(action, resource, agent_id, **ctx):
    if action == "email.send":
        return Decision.REQUIRE_APPROVAL
    return Decision.ALLOW

guard = ExecutionGuard(policy=policy)

@guard.wrap_action(action="email.send", resource="customer@example.com")
def send_email():
    ...

send_email()

Platform operator path

pip install "sqube-agent-guard[control-plane]"
sqube-agent-guard serve --data-dir .sqube

First visit: http://127.0.0.1:8080/setup — create the administrator account in the browser (no terminal password required).

Documentation map

Topic	Page
Concepts	Concepts
Architecture	Architecture
Control plane	Control plane
Event ingestion & SSE	Event ingestion
Sign-in, roles, SSO	Control plane authentication
Policies	Policies
CLI	CLI
Observability	Observability

Source: GitHub