latchgate-openai-agents 0.1.0

OpenAI Agents SDK integration for LatchGate — execution security kernel for AI agents

latchgate-openai-agents

OpenAI Agents SDK integration for LatchGate — execution security kernel for AI agents.

Every tool call goes through LatchGate's enforcement pipeline: auth => policy => WASM sandbox => verification => signed receipt. The LLM never holds credentials and never contacts external systems directly.

Installation

pip install latchgate-openai-agents

Requires a running LatchGate instance:

curl -fsSL https://raw.githubusercontent.com/latchgate-ai/latchgate/main/install.sh | bash && latchgate up

Quick start

from agents import Agent, Runner
from latchgate_openai_agents import latchgate_tools

async def main():
    tools = await latchgate_tools(gate_url="http://localhost:3000")

    agent = Agent(
        name="Secure Worker",
        instructions="You have access to LatchGate-protected tools with full audit trail.",
        tools=tools,
    )

    result = await Runner.run(agent, "Fetch https://httpbin.org/get")
    print(result.final_output)

API

latchgate_tools(**kwargs)

Async factory — discovers actions and returns FunctionTool instances:

tools = await latchgate_tools(
    gate_url="http://localhost:3000",    # Required (or set LATCHGATE_URL)
    agent_id="my-agent",                 # Default: "openai-agents"
    include={"http_fetch", "database"},  # Optional
    exclude={"send_message"},            # Optional
    on_audit=my_audit_callback,          # Optional: receipt callback
)

agent = Agent(name="Worker", tools=tools)

latchgate_tools_from_descriptors(descriptors, *, client)

Create tools from pre-fetched descriptors (no network):

from latchgate_openai_agents import latchgate_tools_from_descriptors, ActionDescriptor

create_tool(descriptor, client)

Create a single FunctionTool from an ActionDescriptor:

from latchgate_openai_agents import create_tool

Error handling

LatchGate errors are returned as structured strings (not raised):

LatchGate error	Tool returns
Policy denied	"ERROR: Action '...' denied: {reason}..."
Approval required	"ERROR: ... requires human approval..." (approval_id emitted via log, not to the model)
Budget exhausted	"ERROR: Budget exhausted..."
Invalid JSON input	"ERROR: Invalid JSON input..."
Transport / infra	"ERROR: LatchGate error..."

Output format

Tool output is a JSON string containing only the action result. Enforcement metadata (receipt ID, trace ID, verification) is never returned to the model — it is emitted at INFO log level and via the optional on_audit callback.

{"status": 200, "body": "{...}"}

License

Apache-2.0