governor-sdk

Python SDK for GovernorAI — the execution boundary for autonomous AI.

Autonomous AI agents can think anywhere — but with GovernorAI they can only act through a policy enforcement gateway. This SDK is how your Python agents talk to it: a @governed decorator, a universal govern() wrapper that auto-detects the framework, and native adapters for the major agent runtimes.

---

Install

pip install governor-sdk

Optional framework adapters:

pip install "governor-sdk[langchain]"     # LangChain + langchain-core
pip install "governor-sdk[langgraph]"     # LangGraph tool nodes
pip install "governor-sdk[llamaindex]"    # LlamaIndex query engines
pip install "governor-sdk[crewai]"        # CrewAI crews and tasks
pip install "governor-sdk[bedrock]"       # AWS Bedrock + boto3
pip install "governor-sdk[gcp]"           # GCP Vertex AI + google-cloud-aiplatform
pip install "governor-sdk[azure]"         # Azure AI Foundry + azure-ai-agents
pip install "governor-sdk[databricks]"    # Databricks + databricks-agents + mlflow
pip install "governor-sdk[all]"           # all of the above

---

Quickstart

import os
from governor import GovernorAIClient, GovernorAIDenied

client = GovernorAIClient(
    base_url=os.environ["GOVERNOR_URL"],
    api_key=os.environ["GOVERNOR_API_KEY"],
)

try:
    result = client.execute(
        tool="erp.process_payment",
        args={"amount": 50_000, "vendor": "ACME"},
        agent_id="finance-agent-v1",
    )
    print("Allowed:", result)
except GovernorAIDenied as e:
    print("Denied by policy:", e)

Or wrap any agent in one line:

from governor import govern

governed = govern(your_langchain_executor)
result = governed.invoke({"input": "process payment for ACME"})

govern() auto-detects LangChain, LangGraph, LlamaIndex, and CrewAI agents and applies the right adapter.

---

Features

• Sync and async clients — GovernorAIClient and AsyncGovernorAIClient
• One-line governance — govern(agent) for any supported framework
• @governed decorator — wrap any function so it routes through a policy check before executing
• Typed decisions — allow / deny / needs-approval surface as GovernorAIDenied and GovernorAIApprovalRequired exceptions
• Fail-closed by default — if the gateway is unreachable, calls deny rather than silently proceeding
• Framework adapters — LangChain, LangGraph, LlamaIndex, CrewAI, AWS Bedrock, Google Cloud ADK / Vertex AI, Azure AI Foundry, Databricks Mosaic AI, AutoGPT
• Session and step tracking — every tool call carries agent_id and session_id so the gateway can enforce per-session step limits and cost caps
• OpenTelemetry-friendly — trace_id propagation built in
• Mypy-strict — type hints throughout

---

Configuration

The client auto-configures from environment variables when explicit values aren't passed:

Variable	Description
GOVERNOR_URL	Gateway URL (default: http://localhost:8080)
GOVERNOR_API_KEY	API key for authentication
GOVERNOR_ORG_ID	Organization ID (sent as X-Org-ID header)

# All three of these are equivalent if env vars are set:
client = GovernorAIClient()
client = GovernorAIClient(base_url=os.environ["GOVERNOR_URL"])
client = GovernorAIClient(base_url="...", api_key="...")

---

Patterns

1. Decorate any function

from governor import governed_fn

@governed_fn("crm.create_case", agent_id="support-agent")
def create_case(customer_id: str) -> dict:
    return {"case_id": "CASE-001", "customer_id": customer_id}

# Governance check happens before the function body runs.
# Raises GovernorAIDenied if the policy denies.
create_case(customer_id="123")

2. LangChain callback handler

from langchain.agents import AgentExecutor
from governor import GovernorAIClient
from governor.integrations.langchain import GovernorAICallbackHandler

client = GovernorAIClient()
callback = GovernorAICallbackHandler(client, agent_id="research-agent")
executor = AgentExecutor(agent=agent, tools=tools, callbacks=[callback])

3. LangGraph tool node

from governor.helpers import govern_langgraph

tool_node = govern_langgraph([search_web, write_file], agent_id="ops-agent")
graph.add_node("tools", tool_node)

4. CrewAI crew

from governor import GovernorAIClient
from governor.integrations.crewai import GovernorAICrew

crew = GovernorAICrew(
    crew=your_crew,
    governor_client=GovernorAIClient(),
    crew_id="research-crew-v2",
)

5. AWS Bedrock agent

import boto3
from governor import GovernorAIClient
from governor.integrations.bedrock import GovernorAIBedrockAgent

agent = GovernorAIBedrockAgent(
    bedrock_client=boto3.client("bedrock-agent-runtime"),
    governor_client=GovernorAIClient(),
    agent_id="bedrock-agent-v1",
)
response = agent.invoke_agent(
    agentId="...",
    agentAliasId="...",
    sessionId="session-123",
    inputText="Process a payment",
)

See examples/quickstart.py for the full set of integration patterns.

---

Decision model

Every call returns one of three outcomes:

Outcome	Python surface
Allow	Returns the gateway response dict containing the tool result
Deny	Raises GovernorAIDenied with policy reason
Needs approval	Raises GovernorAIApprovalRequired with approval URL — the agent can either block, poll, or hand off to a human

from governor import (
    GovernorAIClient,
    GovernorAIDenied,
    GovernorAIApprovalRequired,
    GovernorAIError,
)

try:
    result = client.execute(tool="...", args={...}, agent_id="...")
except GovernorAIDenied as e:
    log.warning("Denied: %s", e)
except GovernorAIApprovalRequired as e:
    notify_human(e.approval_url)
except GovernorAIError as e:
    # gateway unreachable, auth failed, etc — fail closed
    log.error("Gateway error: %s", e)

---

CLI

The SDK installs a small CLI for quick checks:

governorai --help
governorai check tool=erp.process_payment agent_id=finance-agent

---

Development

git clone https://github.com/SentinelLayer/GovernorAI
cd GovernorAI/sdk/python
pip install -e ".[all,dev]"
pytest
ruff check governor
mypy governor

---

End-to-end testing

The default pytest run only exercises unit tests. To verify the SDK against a real gateway, opt into the integration suite under tests/integration/.

Start a gateway locally:

docker-compose up -d gateway control-plane postgres redis

Create an API key (via the control-plane UI at http://localhost:3000 or by POSTing to /api/v1/auth/apikeys) and seed the smoke-test policy into the smoke namespace:

curl -X POST "$GOVERNOR_URL/api/v1/policies/opa" \
  -H "Authorization: Bearer $GOVERNOR_API_KEY" \
  -H "Content-Type: application/json" \
  --data-binary @- <<EOF
{"namespace": "smoke",
 "rego": $(jq -Rs . < tests/integration/policies/smoke_test.rego)}
EOF

Then run the suite:

export GOVERNOR_INTEGRATION=1
export GOVERNOR_URL=http://localhost:8080
export GOVERNOR_API_KEY=gov_xxx
pytest tests/integration/ -v

Or use the wrapper script, which validates the gateway is up and runs pytest for you:

./examples/smoke_test.sh

The integration suite exercises four invariants:

• Allow — smoke.echo is permitted; response shape is well-formed.
• Deny — smoke.export_pii is denied; the SDK raises GovernorAIDenied with a non-empty reason.
• Approval-required — smoke.process_payment requires approval; the SDK raises GovernorAIApprovalRequired.
• Fail-closed — a closed gateway port raises GovernorAIError instead of silently allowing. This is the SDK's most important safety invariant.

Integration tests are auto-skipped when GOVERNOR_INTEGRATION is unset, so the default pytest invocation stays fast and CI-friendly.

---

Compatibility

• Python 3.9, 3.10, 3.11, 3.12
• Works with LangChain ≥1.2, LangGraph ≥1.1, LlamaIndex ≥0.14, CrewAI ≥0.11

---

Links

• PyPI: https://pypi.org/project/governor-sdk/
• Source: https://github.com/SentinelLayer/GovernorAI/tree/main/sdk/python
• Issues: https://github.com/SentinelLayer/GovernorAI/issues
• GovernorAI platform: https://governorai.ai

---

License

Apache License 2.0. See LICENSE.