starvex 1.2.0

Starvex - Production-ready AI agents with guardrails, observability, and security

Starvex

Production-ready AI agents with guardrails, observability, and security.

Installation

pip install starvex

Quick Start

1. Get your API key at starvex.in/dashboard

2. Login with your API key:

starvex login

3. Use in your code:

from starvex import Starvex

vex = Starvex()

async def my_agent(prompt: str) -> str:
    # Your AI agent logic here
    return "Agent response"

# Secure your agent with guardrails
result = await vex.secure(
    prompt="User input",
    agent_function=my_agent
)

if result.status == "success":
    print(result.response)
elif result.status == "blocked":
    print(f"Blocked: {result.verdict.value}")

Features

Guardrails

Automatically detect and block:

• Jailbreak attempts - Prompt injection attacks
• PII leakage - Email, phone, SSN, credit cards
• Toxic content - Harmful or inappropriate text
• Competitor mentions - Block competitor references (configurable)

Observability

Full visibility into your AI agent behavior:

• Real-time event tracing
• Latency monitoring
• Success/failure analytics
• Dashboard at starvex.in/dashboard

Evaluation

Quality checks on agent outputs:

• Hallucination detection
• Faithfulness scoring
• Relevancy analysis

CLI Commands

# Login with your API key
starvex login

# Check connection status
starvex status

# Check a prompt for safety
starvex check "Your prompt here"

# Test a prompt/response pair
starvex test --prompt "Question" --response "Answer"

# Show current configuration
starvex config --show

# Show version
starvex version

# Logout
starvex logout

API Reference

Starvex

from starvex import Starvex

vex = Starvex(
    api_key="sv_live_xxx",  # Or set STARVEX_API_KEY env var
    redact_pii=False,       # Mask PII before logging
    enable_tracing=True,    # Enable observability
)

Methods

secure(prompt, agent_function, context=None)

The main method - secures an AI agent interaction:

1. Checks input safety (jailbreak, PII, toxicity)
2. Calls your agent function
3. Checks output quality (hallucination, toxicity)
4. Logs everything to the dashboard

result = await vex.secure(
    prompt="User input",
    agent_function=my_agent,
    context=["Optional context for hallucination checking"],
    user_id="optional-user-id",
    session_id="optional-session-id",
)

protect(prompt)

Check input only (no agent execution):

result = await vex.protect("User input")
if result.status == "blocked":
    print("Input blocked!")

test(prompt, response, context=None)

Test mode - evaluate a prompt/response pair without tracing:

result = vex.test(
    prompt="What is 2+2?",
    response="2+2 equals 4",
    context=["Math fact: 2+2=4"]
)

Response Object

class GuardResponse:
    status: str          # "success", "blocked", "flagged"
    response: str        # The agent response (or block message)
    verdict: str         # PASSED, BLOCKED_JAILBREAK, BLOCKED_PII, etc.
    checks: list         # Detailed check results
    trace_id: str        # Unique trace ID for debugging
    latency_ms: float    # Total processing time
    warning: str | None  # Warning message if flagged

Environment Variables

Variable	Description
STARVEX_API_KEY	Your Starvex API key
STARVEX_API_HOST	Custom API host (optional)
STARVEX_REDACT_PII	Set to "true" to mask PII
STARVEX_LOG_LEVEL	Logging level (DEBUG, INFO, WARNING, ERROR)

Dashboard

View your metrics and manage API keys at starvex.in/dashboard

• Real-time request monitoring
• Block rate analytics
• Latency tracking
• API key management
• Guardrail configuration

Links

• Website: starvex.in
• Dashboard: starvex.in/dashboard
• Documentation: starvex.in/docs

License

MIT