nim-agent-guardrails • Runtime Guardrails for NVIDIA NIM Agents

Live in the browser

Interactive Guardrails Demo

Test the exact same guardrail logic that runs in Python — no API key, no backend.

View Python source

USER INPUT

MODEL OUTPUT (for output guards)

TOOL CALLS (JSON array)

Results will appear here after running the demo

Quick failing cases

6 built-in primitives

Composable by default

InputLengthGuardrail INPUT

Rejects inputs exceeding char or estimated token limits

Default: 4000 chars / 1500 tokens

BlockedTopicGuardrail IN / OUT

Blocks messages matching configurable regex patterns (weapons, illegal, etc.)

Case-insensitive • Custom patterns supported

PiiDetectionGuardrail OUTPUT (opt-in input)

Detects emails, phones, SSNs, credit cards with regex

Selective PII types • Works on input or output

OutputFormatGuardrail OUTPUT

Enforces max length, required sections, and prohibited phrases

Great for structured responses

CostLimitGuardrail INPUT

Cumulative + per-request token budget enforcement

Stateful across turns • Auto records usage

ToolCallAuditGuardrail TOOL

Allow/block lists, max calls per turn, full audit trail

Audit log always populated

Click any card above to load a matching scenario into the live demo →

Simple & composable

Wrap every turn in safety

Guardrails run before the NIM call, around tool execution, and on the final output. Short-circuit on the first failure.

1. INPUT GUARDRAILS

Length + token estimate

Blocked topics / jailbreaks

Budget check (pre-flight)

2. NVIDIA NIM CALL

Uses the official OpenAI-compatible endpoint at integrate.api.nvidia.com/v1.
Full tool calling support. Automatic usage tracking.

Nemotron • Llama 3.1 • Any NIM model

3. OUTPUT + TOOL GUARDS

Tool allow/block + rate limits

PII redaction / blocking

Format + prohibited phrases

Every turn produces a structured TurnRecord + full audit log (JSON).

Drop-in usage

Code that feels native

from nim_guardrails import (
    NimClient, GuardedAgent, AgentConfig,
    GuardrailChain, InputLengthGuardrail,
    BlockedTopicGuardrail, PiiDetectionGuardrail, CostLimitGuardrail
)

client = NimClient()   # reads NVIDIA_API_KEY

input_guards = GuardrailChain([
    InputLengthGuardrail(max_chars=2000),
    BlockedTopicGuardrail(),
    CostLimitGuardrail(max_total_tokens=50_000),
])

agent = GuardedAgent(
    client=client,
    config=AgentConfig(system_prompt="You are a helpful travel assistant."),
    input_guardrails=input_guards,
    output_guardrails=GuardrailChain([PiiDetectionGuardrail()]),
)

response, record = agent.run("Plan a trip to Tokyo")
print(response)
print("Blocked?", record.blocked)
print("Latency:", record.latency_ms)

from nim_guardrails import Guardrail, GuardrailResult

class LanguageGuardrail(Guardrail):
    name = "language_check"

    def check(self, *, output_text=None, **_):
        if output_text and "sorry" in output_text.lower():
            return self._fail("Model apologised — likely refusing")
        return self._ok("No refusal language")

# Add to any chain
output_guardrails.add(LanguageGuardrail())

# Full structured audit after any run
import json
print(json.dumps(agent.get_audit_log(), indent=2))

# Example record:
# {
#   "turn": 1, "blocked": false, "latency_ms": 1240,
#   "usage": {"total_tokens": 231},
#   "guardrails": {
#     "input":  [{"name": "input_length", "passed": true, ...}],
#     "output": [{"name": "pii_detection", "passed": true, ...}],
#     "tool":   []
#   }
# }

All examples run with pip install "git+https://...[nim]"

Guardrails thatactually run in-process.

Interactive Guardrails Demo

Composable by default

Wrap every turn in safety

Code that feels native

Guardrails that
actually run
in-process.