Log every LLM call, tool use, and action your agents take. Get alerted when something looks wrong — before your users do.
No credit card required · Free tier available
Every LLM call, tool execution, and agent action logged in real time. REST API or one-line SDK — Node.js and Python both supported.
11 detection rules across prompt injection, data exfiltration, PII exposure, and secret leakage. Scores are probabilistic — no alert storms.
High and critical events surface as in-app alerts with Slack and webhook delivery. Configurable threshold so you tune sensitivity, not noise.
Organization isolation with role-based access (admin, member, viewer). Invite teammates, switch orgs, manage API keys per environment.
Add the SDK to your agent. One line of code. Works with LangChain, OpenAI, Anthropic, or raw HTTP.
Argus captures every event, scores risk, and chains tamper-evident audit hashes in real time.
Dismiss alerts, fix the agent, or export the immutable event chain to share with your security team or auditor.
Your event data is yours — retained for your configured period, exportable at any time, never used to train models.
Drop the SDK into your existing agent code. Works with LangChain, OpenAI, Anthropic, and raw HTTP calls.
import { ArgusClient } from "argus-sdk";
const argus = new ArgusClient({
apiKey: "sk_live_...",
actorId: "my-agent",
});
await argus.llmCall({
inputText: "Analyze this contract",
outputText: "The contract includes...",
inputTokens: 1024,
outputTokens: 256,
});
// Risk scored + hashed in < 50msAny agent that makes LLM calls or takes tool/action steps — LangChain, OpenAI Agents SDK, Anthropic Claude, CrewAI, AutoGen, or custom pipelines. If it can make an HTTP request, it can send events to Argus.
Argus applies 11 probabilistic detection rules across four categories: input/output content, action type, token usage, and workflow patterns. Scores are combined using 1−∏(1−wᵢ) to prevent saturation — so many small signals don't automatically create a false critical alert.
When multiple correlated events (sharing a correlationId) show increasing risk scores over the course of a single agent run, Argus flags it as an escalating workflow. This catches multi-step attacks that no single event would trigger on its own.
Yes — input and output text are stored to power risk analysis and audit review. You can enable PII redaction on ingestion, set a retention period (default 90 days), and trigger GDPR erasure via the API.
Risk scoring runs synchronously at ingestion in under 50ms. Alert creation and Slack/webhook delivery happen asynchronously and typically complete within a few seconds.
Yes. The settings page lets you define custom keywords (e.g. wire_transfer, ssn_lookup). Any event whose input or output contains a match gets an additional risk score boost. You can also tune the global alert threshold.
Yes — the free plan includes 10,000 events per month, 30-day retention, and 1 seat. No credit card required to get started.
Install the SDK (npm install argus-agent-sdk or pip install argus-agent-sdk), create an API key in the dashboard, and wrap your LLM calls. Most integrations take under 10 minutes.
Get started in minutes. No infrastructure to manage.
Get started free →