Trace AI Agent Activity with Trident SDK

Tracing gives you a complete audit trail of everything your agent does: every prompt it receives, every LLM call it makes, every tool it invokes, and every response it generates. Without a trace, a misbehaving agent is a black box — you know something went wrong, but not where, why, or what the model actually said. Trident’s tracing layer captures the full picture in production so you can investigate incidents, verify firewall decisions, and feed accurate data into red-team replay.

How tracing works

The Trident SDK wraps OpenLLMetry (the open-source OpenTelemetry instrumentation layer for LLMs, built by Traceloop). When you call trident.init(), the SDK patches every LLM client loaded in the same process — OpenAI, Anthropic, LangChain, and others — so each LLM call automatically becomes an OTEL span. Those spans are shipped to your Trident project’s OTLP ingest endpoint (/api/public/otel/v1/traces) using Basic auth derived from your project key pair. You do not need to wrap individual calls or modify your agent logic. One init() call at startup covers everything in the process.

Installation and setup

TypeScript
Python

Install the SDK from npm:

npm install @vouch-ai/sdk

Call trident.init() once at application startup, before any LLM calls:

import { trident } from "@vouch-ai/sdk";

trident.init({
  projectPk: process.env.TRIDENT_PROJECT_PUBLIC_KEY,
  projectSk: process.env.TRIDENT_PROJECT_SECRET_KEY,
  agentId: "prod-rag-bot", // optional — scopes spans to this agent
});

// Any LLM call after init() is automatically traced.
import OpenAI from "openai";
const openai = new OpenAI();
await openai.chat.completions.create({ model: "gpt-4o", messages: [...] });

Install the SDK from PyPI:

pip install vouch-sdk

Call vouch_sdk.init() once at application startup:

import vouch_sdk

vouch_sdk.init(
    project_pk="pk-...",
    project_sk="sk-...",
    agent_id="prod-rag-bot",  # optional — scopes spans to this agent
)

# Any LLM call after init() is automatically traced.
from openai import OpenAI
client = OpenAI()
client.chat.completions.create(model="gpt-4o", messages=[...])

Environment variables

Instead of passing keys directly to init(), you can set environment variables and call init() with no arguments:

TRIDENT_PROJECT_PUBLIC_KEY=pk-...
TRIDENT_PROJECT_SECRET_KEY=sk-...
TRIDENT_AGENT_ID=prod-rag-bot       # optional

The TypeScript and Python SDKs also read VOUCH_PROJECT_PUBLIC_KEY and VOUCH_PROJECT_SECRET_KEY as back-compat aliases.

Zero-code option for Node.js

If you want to add tracing to an existing Node.js application without modifying its source code, use the --require flag to load the SDK’s auto-instrumentation register at process start:

node --require @vouch-ai/sdk/register server.js

Set your TRIDENT_PROJECT_PUBLIC_KEY and TRIDENT_PROJECT_SECRET_KEY environment variables before running the command. No code changes required.

What each trace contains

Every traced LLM call produces a span with the following attributes:

Attribute	Description
Prompt text	The full messages array sent to the model
Model name	The model identifier (e.g. `gpt-4o`, `claude-3-5-sonnet`)
Tool calls	Any function/tool calls the model made, including arguments
MCP invocations	Model Context Protocol tool calls and their results
Response	The full model response text
Latency	End-to-end call duration in milliseconds
Token counts	Prompt tokens, completion tokens, and total

Spans are grouped into traces by request, so a single multi-step agent run — retrieval, reasoning, tool call, summarization — appears as one coherent trace tree in the dashboard.

PII redaction

The SDK scrubs sensitive data from span attributes inside your process, before any trace data is transmitted to Trident. The following patterns are redacted and replaced with a labelled placeholder (e.g. [EMAIL], [CREDIT_CARD]):

Email addresses — user@example.com
Credit card numbers — validated with Luhn checksum to avoid false positives
US Social Security Numbers — 123-45-6789
AWS access key IDs — AKIA..., ASIA..., AIDA...
JSON Web Tokens — three-part eyJ... strings
API keys — sk- and sk-ant- prefixed secrets
IBANs — international bank account numbers
IPv4 addresses
Phone numbers

Redaction is enabled by default. To supply your own rules, pass a redactPII option to init():

TypeScript
Python

import { trident } from "@vouch-ai/sdk";
import type { PiiRule } from "@vouch-ai/sdk";

const myRules: PiiRule[] = [
  { name: "EMPLOYEE_ID", pattern: /EMP-\d{6}/g },
];

trident.init({
  projectPk: process.env.TRIDENT_PROJECT_PUBLIC_KEY,
  projectSk: process.env.TRIDENT_PROJECT_SECRET_KEY,
  redactPII: { rules: myRules }, // replaces the built-in rule set
});

import vouch_sdk
from vouch_sdk import PiiRule
import re

my_rules = [
    PiiRule("EMPLOYEE_ID", re.compile(r"EMP-\d{6}")),
]

vouch_sdk.init(
    project_pk="pk-...",
    project_sk="sk-...",
    redact_pii={"rules": my_rules},  # replaces the built-in rule set
)

To disable PII redaction entirely, pass redactPII: false (TypeScript) or redact_pii=False (Python). Only do this if you have verified that your traces contain no sensitive data.

Viewing traces

Open the Traces tab in the Trident dashboard to see all captured spans. You can filter by:

Agent — scope the view to a specific agentId
Time range — narrow to a deployment window or incident window
Model — compare behaviour across model versions
Trace ID — jump directly to the span linked from a finding

Select any trace to see the full span tree, including prompt text (post-redaction), model responses, tool call arguments, and latency breakdown.

Tracing overhead is sub-millisecond. The SDK uses OpenTelemetry’s BatchSpanProcessor by default, which buffers spans and exports them asynchronously. Your agent’s request latency is not measurably affected.

​How tracing works

​Installation and setup

​Environment variables

​Zero-code option for Node.js

​What each trace contains

​PII redaction

​Viewing traces

How tracing works

Installation and setup

Environment variables

Zero-code option for Node.js

What each trace contains

PII redaction

Viewing traces