cortexhub 0.5.25

CortexHub Python SDK — Runtime governance for AI systems. One decorator, any tool, any framework.

CortexHub Python SDK

Runtime governance for AI systems. One decorator. Any tool. Any framework.

pip install cortexhub

Python 3.10–3.12.

Optional installs (match your agent stack)

Use bracket extras so peer dependencies install in one step — same shape as the demo app:

Extra	Typical use	Pair with cortexhub.init(framework=...)
openai-agents or openai-agent	OpenAI Agents SDK	openai_agents
langchain	LangChain / shared ChatModel	langchain
langgraph	LangGraph + OpenAI chat models (pulls langgraph, langchain-openai, …)	langgraph
anthropic or claude	Claude / Anthropic Messages API	anthropic, claude, or claude_agents
crewai, llama-index, autogen, pydantic-ai, semantic-kernel, google-adk	Those ecosystems	Fleet label only* — use cx.llm_call for LLM spans unless you route via LangChain / Anthropic / OpenAI hooks
agents	LangGraph + OpenAI Agents together (demo-style)	Set framework= per vertical

* Automatic llm.call hooks today: OpenAI Agents, LangChain / LangGraph (BaseChatModel.invoke / ainvoke), Anthropic (messages.create, non-streaming). Everything else uses the same @cx.tool policy / approval behaviour; wrap LLM calls with cx.llm_call / cx.record_llm_result or adopt a hooked client above.

pip install 'cortexhub[langgraph,openai-agents]'

Quick start

Identity comes from your agent key only — the control plane assigns the canonical agent id. Set CORTEXHUB_AGENT_KEY (or pass agent_key= to init()).

import cortexhub

cx = cortexhub.init()  # requires CORTEXHUB_AGENT_KEY (or pass agent_key=)

@cx.tool(classification="write")
def process_payment(customer_id: str, amount: str, reference: str) -> dict:
    """Process a payment."""
    return payment_service.charge(customer_id, amount, reference)

# Policies, approvals, and audit trail are configured in the control plane.

Configuration

Env var	Description
CORTEXHUB_AGENT_KEY	Agent identity secret (required for cloud enforcement)
CORTEXHUB_DEBUG	Set to 1 or true for readable SDK logs: tool + arguments, policy name + condition, and decision
CORTEXHUB_LOG_LEVEL	e.g. DEBUG, INFO (default INFO)

Debug logging

With CORTEXHUB_DEBUG=1 (or CORTEXHUB_DEBUG_POLICY=1), the SDK logs in real time:

• Tool encountered — tool name and argument names/values (values redacted when privacy mode is on)
• Policy evaluated — policy name, condition/reasoning, and decision (allow / deny / escalate)
• Decision — final effect, policy, and reason for the current tool call

Useful for seeing exactly what CortexHub is evaluating and why a tool was allowed, blocked, or sent for approval.

OpenAI Agents SDK (llm.call telemetry)

If you use the OpenAI Agents SDK, initialise with framework="openai_agents" so each model turn is captured as llm.call spans (assistant reasoning text and tool-call intent), merged automatically with any hooks= you pass to Runner.run.

pip install 'cortexhub[openai-agents]'
# or: pip install 'cortexhub[openai-agent]'

cx = cortexhub.init(framework="openai_agents")

---

Documentation · Control plane