Decision OS Gate

Integrations

Gate is a plain HTTP API -- there's nothing to install to start calling it from any language. Get your key from your account page first, then call verify-action right before your agent actually executes a tool call.

The API, directly (any language)

One call, before the tool runs:

curl -X POST https://gate.inmem.ai/v1/verify-action \
  -H "Authorization: Bearer <YOUR_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "please cancel order #W123456"}],
    "tool_call": {"name": "cancel_order", "arguments": {"order_id": "#W123456"}}
  }'

Returns {"approved": true/false, "checks": [...]} -- if approved is false, don't run the tool call; show the agent the checks that failed so it can recover or ask a human.

Python

Stdlib only, no extra package required (an official pip-installable SDK is on the way):

import json, urllib.request

def verify_action(messages, tool_call, api_key):
    req = urllib.request.Request(
        "https://gate.inmem.ai/v1/verify-action",
        data=json.dumps({"messages": messages, "tool_call": tool_call}).encode(),
        headers={
            "Authorization": f"Bearer {api_key}",
            "Content-Type": "application/json",
        },
        method="POST",
    )
    with urllib.request.urlopen(req) as resp:
        return json.load(resp)

result = verify_action(
    messages=[{"role": "user", "content": "please cancel order #W123456"}],
    tool_call={"name": "cancel_order", "arguments": {"order_id": "#W123456"}},
    api_key="<YOUR_API_KEY>",
)
if not result["approved"]:
    ...  # don't execute the tool call -- see result["checks"] for why

Node.js / TypeScript

const resp = await fetch("https://gate.inmem.ai/v1/verify-action", {
  method: "POST",
  headers: {
    Authorization: "Bearer <YOUR_API_KEY>",
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    messages: [{ role: "user", content: "please cancel order #W123456" }],
    tool_call: { name: "cancel_order", arguments: { order_id: "#W123456" } },
  }),
});
const result = await resp.json();
if (!result.approved) {
  // don't execute the tool call -- see result.checks for why
}

Agent frameworks (LangChain, LangGraph, custom tool-calling loops)

Wrap whatever already executes your tool call, call verify-actionfirst, and raise/refuse if it comes back unapproved -- that works the same whether you're on LangChain's BaseTool, LangGraph, AutoGen, CrewAI, or a hand-rolled loop, since it only touches the point where a tool call is about to run, not any one framework's internals.

No addons installed yet? Your account starts with an empty protocol set on purpose -- install a ready-made domain-task + platform addon (or confirm your own tool names) from your account page before verify-action has anything to check against.