Gateplex Python SDK

Real-time AI governance firewall for LLM applications. Drop-in intercept logging, PII detection, spend controls, and human-in-the-loop approval workflows — deployable inside your VPC.

Installation

pip install gateplex-python

Quick start

from gateplex import GateplexClient

client = GateplexClient(api_key="gp-...")

response = client.log_intercept(
    agent_id="<your-agent-uuid>",
    event_type="llm_call",
    input="Summarise the patient record for John Smith",
    output="John Smith, DOB 1982-03-14, diagnosis: ...",
)

if response.is_blocked:
    raise RuntimeError("Request blocked by governance policy")

print(response.verdict, response.pii_types_detected)

Verdict values

log_intercept automatically raises GateplexBlockedError on a BLOCK verdict.

Human-in-the-loop approvals

from gateplex import GateplexClient, GateplexRejectedError

client = GateplexClient(api_key="gp-...")

try:
    response = client.log_intercept(
        agent_id="<agent-uuid>",
        event_type="llm_call",
        input=user_prompt,
        output=llm_output,
        wait_for_approval=True,   # blocks until reviewer acts
        poll_timeout=600,         # seconds (default 300)
    )
except GateplexRejectedError as e:
    print("Reviewer rejected the request:", e)

You can also poll manually:

response = client.log_intercept(...)

if response.is_pending_approval:
    approval = client.wait_for_approval(response.intercept_id)
    print(approval.status, approval.reviewer_note)

InterceptResponse fields (v0.2.3)

response.intercept_id        # str
response.verdict             # "ALLOW" | "FLAG" | "BLOCK" | "PENDING_APPROVAL"
response.reasoning           # str — human-readable explanation
response.triggered_rules     # List[str] — rule IDs/names that fired
response.pii_types_detected  # List[str] — e.g. ["EMAIL", "SSN"]
response.transaction_amount  # Optional[float]
response.requires_approval   # bool

# Convenience properties
response.is_allowed
response.is_flagged
response.is_blocked
response.is_pending_approval

ApprovalStatus fields (v0.2.3)

approval.intercept_id   # str
approval.status         # "PENDING" | "APPROVED" | "REJECTED"
approval.reviewed_by    # Optional[str]
approval.reviewed_at    # Optional[str] — ISO 8601
approval.reviewer_note  # Optional[str]

approval.is_approved
approval.is_rejected
approval.is_pending

Background batch logging

# Fire-and-forget — does not block the hot path
client.batch_intercept(
    agent_id="<agent-uuid>",
    event_type="llm_call",
    input=prompt,
    output=completion,
)

# Drain the queue before process exit
client.flush()

LLM library integrations

from gateplex.integrations import OpenAIPatcher, AnthropicPatcher
import openai

openai_client = openai.OpenAI(api_key="sk-...")
patcher = OpenAIPatcher(client, agent_id="<agent-uuid>", openai_client=openai_client)
patcher.patch()
# All openai_client.chat.completions.create calls are now logged automatically
patcher.unpatch()

@monitor decorator

from gateplex.integrations import monitor

@monitor(client, agent_id="<agent-uuid>", event_type="llm_call")
def call_llm(input: str) -> str:
    return my_llm(input)

Exceptions

Context manager

with GateplexClient(api_key="gp-...") as client:
    client.log_intercept(...)

Changelog

0.2.3

• InterceptResponse gains verdict, reasoning, triggered_rules, pii_types_detected, transaction_amount, requires_approval, and is_* convenience properties.
• New ApprovalStatus dataclass.
• GateplexClient.wait_for_approval() — polls until approval resolves.
• log_intercept gains wait_for_approval parameter.
• New GateplexBlockedError and GateplexRejectedError exceptions.
• Guardrail gains requires_approval field.

0.2.2

• @monitor decorator and OpenAIPatcher / AnthropicPatcher integrations.

0.2.1

• Initial public release.