AI guardrails that actually hold

MSPClaw uses NemoClaw — NVIDIA's enterprise AI safety framework — to enforce input/output rails, block prompt injection, and keep every interaction inside your defined operational boundaries. Guardrails aren't a feature. They're the foundation.

What is NemoClaw?

NemoClaw is NVIDIA's open-source guardrails framework built on NeMo Guardrails. It lets you define conversational rails in Colang — specifying exactly what an AI can and cannot say or do — and enforces them at every model call. MSPClaw ships with a purpose-built MSP rail set: tenant isolation, scope enforcement, destructive-action blocks, and compliance-ready audit hooks, all powered by the same framework NVIDIA deploys in enterprise production environments.

NVIDIA NemoClaw ↗

🛡️

Input Rails

Every prompt is checked before it reaches the model. NemoClaw inspects for prompt injection, off-topic requests, and out-of-scope tenant references — refusing or redirecting before the LLM ever processes the input.

🔍

Output Rails

Model responses are validated before delivery. Sensitive data patterns (PII, credentials, internal IPs) are filtered, and any response that would trigger a policy violation is blocked and logged — never surfaced to the tech.

🧭

Topical Rails

MSPClaw only does what it's supposed to do. NemoClaw topical rails prevent the AI from drifting into general-purpose tasks, off-brand responses, or anything outside your defined MSP operational scope.

🚫

Jailbreak Detection

Colang-defined safety rails detect and neutralize adversarial prompts, role-play overrides, and social engineering attempts — so a clever technician (or a compromised one) can't talk the AI into bypassing your controls.

Every request flows through NemoClaw's rail stack

💬

Tech Request (Teams / Slack / Ticket)

→

🛡️

Input Rail Check (NemoClaw)

→

🧭

Topical & Scope Filter

→

🤖

LLM + Tool Execution

→

🔍

Output Rail Check (NemoClaw)

→

👥

Approval Gate (if required)

→

✅

Audited Action + Response

MSP-specific guardrails on top

NemoClaw provides the AI safety layer. MSPClaw adds the MSP-specific operational controls on top.

Approval Gates Prevent unauthorized changes

🛑

Destructive Action Blocks NemoClaw rails block destructive actions (device wipes, user deletions, firewall changes) before they reach the tool layer. A human approval gate is the only path forward.

👥

Role-Based Thresholds Colang rail definitions encode your tier structure: L1 techs operate within a narrow rail set; L3 escalation widens scope under defined conditions.

Tenant Boundaries Hard-scoped data isolation

🔒

Cross-Tenant Prompt Blocks NemoClaw input rails hard-scope every session to a single tenant context. A prompt referencing Company A cannot access, mention, or act on Company B data — at the rail level, not just the app level.

📝

Rail-Triggered Audit Trails Every rail trigger — block, redirect, or pass — is logged with the full prompt context, the matched rule, the outcome, and the requesting user. Compliance-ready by default.

Scope Enforcement Least-privilege AI

🎯

Action Allow-Lists NemoClaw topical rails define an explicit allow-list of permitted MSP operations. Anything outside the list — regardless of how the prompt is worded — is blocked at the rail layer.

⚡

Evidence-Gated Execution High-risk actions require the AI to surface supporting evidence first. NemoClaw output rails verify evidence is present before the action tool is called.

Secure enough for your most paranoid client.

Book a deep-dive demo to walk through our NemoClaw rail configuration and MSP compliance posture.

We'll share our full NemoClaw Colang rail definitions on request.

Book a demo