Skip to main content
This quickstart evaluates your agent against Vijil’s Trust Score harness. You’ll wrap your existing agent function, run an evaluation, and see results—all in about 15 minutes.

​
Prerequisites

  • Python 3.9+
  • A Vijil API key (get one here)
  • An OpenAI API key (or another LLM provider)

​
Install

pip install vijil

​
Set Credentials

export VIJIL_API_KEY="your-vijil-key"
export OPENAI_API_KEY="your-openai-key"

​
Evaluate Your Agent

Create a file called evaluate.py: 
from vijil import Vijil
from openai import OpenAI

# Define your agent as a function
def my_agent(prompt: str) -> str:
    client = OpenAI()
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": prompt}
        ]
    )
    return response.choices[0].message.content

# Evaluate it
vijil = Vijil()

local_agent = vijil.local_agents.create(
    agent_function=my_agent,
    agent_name="my-first-agent"
)

vijil.local_agents.evaluate(
    agent_name="my-first-agent",
    harnesses=["trust_score"]
)
Run it: 
python evaluate.py
The evaluation takes 10–15 minutes. When complete, you’ll see output like: 
Trust Score: 0.82
├── Reliability: 0.91
├── Security: 0.74
└── Safety: 0.85

High-severity findings: 2
Medium-severity findings: 5

​
What Just Happened?

  1. Vijil wrapped your agent in a temporary HTTP server using ngrok
  2. Diamond sent probes — adversarial prompts testing for hallucinations, prompt injection, jailbreaks, and more
  3. Detectors analyzed responses — checking if your agent leaked data, followed malicious instructions, or violated safety policies
  4. Results aggregated into a Trust Score with specific findings
Your agent code wasn’t modified. The evaluation ran against your actual implementation.

​
View Detailed Results

Open the Vijil Console to see:
  • Per-probe results with the exact prompts and responses
  • Failure explanations with remediation guidance
  • Comparison with previous evaluations

​
What’s Next?

Framework Guides

Integrate with LangChain, Google ADK, or custom frameworks

Add Protection

Block attacks at runtime with Dome guardrails

CI/CD Integration

Run evaluations on every pull request

Understanding Results

Interpret scores and prioritize fixes
Last modified on March 19, 2026