ASCEND AI SDK

Official Python SDK for Ascend AI Governance Platform — enterprise AI governance and enforcement for regulated industries.

Govern what your AI agents do — in real time. ASCEND intercepts every agent action, scores risk using CVSS v3.1/NIST 800-30/MITRE ATT&CK, enforces your policies, and maintains an immutable audit trail for compliance.

Built for financial services, healthcare, and government contractors.

Install

pip install ascend-ai-sdk

Quick Start

from ascend import AscendClient, FailMode

client = AscendClient(
    api_key="your_api_key",
    api_url="https://pilot.owkai.app",
    agent_id="my-agent",
    agent_name="My Agent",
    fail_mode=FailMode.CLOSED
)

# Govern an action before it executes
result = client.evaluate_action(
    action_type="database_write",
    resource="customer_db",
    wait_for_decision=False
)

print(result.decision)      # APPROVED / PENDING / DENIED
print(result.risk_score)    # 0-100
print(result.cvss_score)    # CVSS v3.1 base score
print(result.mitre_tactic)  # MITRE ATT&CK tactic

Enforcement Decision Attribution

Inspect why the platform reached its verdict. These fields ship in AuthorizationDecision as of SDK 2.7.0:

# Authoritative governance verdict
print(result.enforcement_decision)
# 'auto_approved' | 'pending_approval' | 'denied' | 'escalated'

# Which signal drove the verdict
print(result.enforcement_decision_source)
# 'threshold' | 'policy' | 'smart_rule' |
# 'code_analysis' | 'prompt_security'

# Which signal contributed the highest risk score
print(result.risk_score_source)
# 'cvss' | 'policy' | 'code_analysis' |
# 'prompt_security' | 'pipeline'

# Shadow scoring — what the system WOULD have decided under
# the org's shadow threshold config. None when no shadow
# config exists (opt-in feature, observational only).
print(result.shadow_enforcement_decision)
print(result.shadow_enforcement_decision_source)

Model Governance (SR-11-7 / EU AI Act Art. 9)

Enforce your model registry at action submit time. Non-compliant or unregistered models are denied.

result = client.evaluate_action(
    action_type="model_inference",
    resource="ml_pipeline",
    model_id="gpt-4-production",  # checked against registry
    wait_for_decision=False
)

print(result.model_governance["registry_checked"])
print(result.model_governance["compliance_status"])

MCP Layer 13 Governance

Govern actions from MCP servers. Unregistered or deactivated servers are denied.

result = client.evaluate_action(
    action_type="tool_call",
    resource="crm_system",
    mcp_server_name="salesforce-mcp",
    wait_for_decision=False
)

print(result.mcp_governance["server_registered"])

Kill-Switch

Block all agent actions within one poll cycle (default 5 seconds). Kill-switch server handler p99=17.03ms (measured June 2, 2026).

# Agent side — poll for kill-switch signals
client.start_kill_switch_polling(
    interval_seconds=5
)

# Fail-secure: agents fail-closed after 3 consecutive unreachable
# polls (~15 seconds worst-case). Recovers automatically when the
# endpoint becomes healthy.
if client.is_blocked():
    # Kill-switch active OR polling has failed 3+ times —
    # do not proceed with the agent action.
    pass

Capabilities

Compliance

SOC 2 Type II · PCI-DSS · HIPAA · FedRAMP Compatible NIST AI RMF · SR 11-7 · EU AI Act Art. 9/28

Links

• Platform: https://pilot.owkai.app
• Documentation: https://docs.ascendowkai.com
• Status: https://ascend-status.instatus.com
• Support: info@ow-kai.com

Requirements

Python 3.8+

ASCEND is a product of OW-KAI Technologies, Inc. 9+ years spanning AI/ML governance and cybersecurity.