Runtime agent observability

Scouter Documentation

Scouter adds runtime semantic authorization to AI agents. Declare what the agent is allowed to do, wrap the client, and evaluate tool calls, writes, and external requests before they execute.

Try the Product Request SDK access Request access

Instrument actions

Capture agent activity at the point where requests become tool calls, records, files, or API requests.

Score risk

Compare each action with declared intent, permissions, data exposure, and runtime context.

Enforce policy

Run in audit mode first, then move to warnings, escalations, or hard stops for unsafe behavior.

Architecture

Scouter sits between the agent framework and the action surface. It triages whether an action needs analysis, estimates consequence, applies specialized guards, and returns an allow, warn, or block decision.

Step 1 User request

Step 2 Action triage

Step 3 Consequence engine

Step 4 Execution guards

Step 5 Allow, warn, or block

Designed for production agents

The SDK supports audit and enforcement workflows so teams can observe behavior, tune policy, and stop high-risk execution without rebuilding the agent from scratch.

Install

Use the private package for your agent runtime. Python supports OpenAI, LangChain, CrewAI, AutoGen, and PhiData. Node supports OpenAI and the Vercel AI SDK.

Python

pip install scouter-ai

Node

npm install @scouter-ai/node

Python quickstart

Create a Scouter client, register the intent, and wrap the OpenAI client before agent calls are made. Start with audit mode when you want visibility first.

python

from scouter.client import ScouterClient
from scouter.integrations.openai import wrap_openai
from openai import OpenAI

scouter = ScouterClient(api_key="scouter_api_key", mode="audit")
intent = scouter.register_intent(
    name="support-refund-agent",
    description="Answer refund questions and create approved support actions.",
)

client = wrap_openai(
    OpenAI(api_key="openai_api_key"),
    scouter=scouter,
    intent_id=intent.intent_id,
)

Node quickstart

The Node SDK follows the same pattern: configure Scouter, describe the agent intent, then wrap the client used by your application.

typescript

import OpenAI from "openai";
import { ScouterClient } from "@scouter-ai/node";
import { wrapOpenAI } from "@scouter-ai/node/integrations/openai";

const scouter = new ScouterClient({
  apiKey: "scouter_api_key",
  mode: "audit",
});

const intent = await scouter.registerIntent({
  name: "support-refund-agent",
  description: "Answer refund questions and create approved support actions.",
});

const client = wrapOpenAI(new OpenAI(), {
  scouter,
  intentId: intent.intentId,
});

Execution guards

Guards inspect specific action types before execution. Use them to identify destructive commands, risky database operations, unsafe API calls, and exposure patterns that policy alone may miss.

Guard	What it watches	Examples
ShellGuard	Terminal and filesystem commands that could damage systems or reveal sensitive files.	Recursive deletes, broad permission changes, access to protected files.
DatabaseGuard	SQL and data-layer actions that modify or expose production data.	Drops, unscoped deletes, suspicious injection patterns.
APIGuard	External requests, credential handling, and high-risk endpoint access.	Secrets in URLs, privileged updates, data exfiltration paths.

Configuration

Connect the SDK to the hosted Scouter backend and dashboard. Create an API key from the Scouter console, then pass it through environment variables or your secret manager.

.env

SCOUTER_BACKEND_URL=http://127.0.0.1:4173
SCOUTER_API_KEY=your-api-key-here

Examples

The private SDK package includes examples for quickstarts, inline guards, OpenAI governance, backend connection, chatbot flows, and guard demonstrations.

Python examples

Quickstart scripts, OpenAI governance, backend connection, and guard demos.

Private SDK package

Review package docs, examples, tests, and release notes once access is provisioned.

Development

Use editable installs for local Python work and the standard Node build scripts for the TypeScript package.

Python package

cd python
pip install -e ".[dev]"
pytest tests/

Node package

cd node
npm install
npm run build
npm test