NVIDIA Secures the Agent. Revenium Governs Its Spend and Proves Its ROI.

Revenium's OpenClaw skill now runs inside the NemoClaw and OpenShell sandbox, adding metering, budget guardrails, and job-level outcome and ROI tracking without weakening NVIDIA's security model.

NemoClaw governs what an always-on agent is allowed to do, but it does not measure what the agent spends or whether that spend is worth it. In a sandbox, costs can quietly compound across completions, tool calls, retries, and subagent fan-out, and ROI is hard to reconstruct after the fact.

Revenium closes that operating gap by showing cost per outcome for agents running on Nemotron, with controls that can warn, block, or run in shadow mode before spend gets away from you.

A quick primer on NemoClaw

Plenty of teams know OpenClaw but have not met NemoClaw yet, so it helps to pull the layers apart, because people tend to run them together.

OpenShell is the hardened sandbox runtime, part of the NVIDIA Agent Toolkit. It provides out-of-process policy enforcement, isolated execution, granular permissions, a privacy router, and a managed egress proxy.

NemoClaw is the orchestration and security layer that deploys and supervises agents inside OpenShell. Its CLI is nemoclaw, and the product page covers the broader platform.

The agent itself, OpenClaw or Hermes, runs inside the sandbox, with inference routed to open models like NVIDIA Nemotron, including Nemotron 3 Super. It can run locally, on-prem, or in the cloud on DGX Spark, DGX Station, or a workstation with an NVIDIA RTX GPU. NVIDIA's own developer guide walks through the setup.

The one-line takeaway is that NemoClaw is about running powerful autonomous agents you can trust, locked down by default.

The operating gap NemoClaw doesn’t try to solve

NemoClaw's guardrails focus on safety and privacy. They define what the agent can execute, what data can move, and which destinations are allowed through egress. Those controls are essential, but they do not measure cost, prevent budget drift, or connect agent work to business value.

Always-on agents make that gap more visible. Spend can build through repeated completions, tool calls, retries, and subagent fan-out. The operator may see a secure agent doing permitted work while the economics remain opaque.

The sandbox also raises the bar for any add-on. Egress is allowlisted, the filesystem is isolated, and HOME is /sandbox. Cost and value governance has to fit that model instead of working around it.

What the Revenium skill does

Revenium for OpenClaw now ships a first-class NemoClaw and OpenShell install path.

Inside the sandbox, Revenium delivers per-completion and tool-event metering, agentic job lifecycle tracking, Revenium Cost Controls, budget guardrails with a warn band, a hard block band, shadow mode, and the foundation for outcome and ROI analysis.

Existing users don’t need to change their current path. The standalone OpenClaw and Docker installation remains separate. The NemoClaw skill path is additive, driven by environment variables, idempotent, and validated on a live host from one command.

Built to respect the sandbox

Cost governance inside a security-first sandbox only works if it keeps the sandbox intact. Revenium follows three rules in the NemoClaw path.

• It plays by the sandbox's rules. Outbound metering goes through NemoClaw's own managed egress policy, so no holes get punched in the proxy.
• It keeps no secrets in the clear. Credentials are delivered into the sandbox and stored in a locked-down, permission-restricted config, never passed on a command line.
• It runs no rogue daemon. Metering runs as a lightweight host-side loop against the sandbox, so nothing extra has to live and run inside it.

The install is open source, including the scripts and runbook. The repository is available on GitHub for review and contribution.

From tokens to outcomes to ROI

The skill connects AI spend to business ROI, so teams can see the business value returned for every dollar spent on AI.

It does that in four stages: meter the spend, track the job lifecycle, attach the business outcome, and calculate ROI.

1. Meter everything with meaning. Every completion, tool call, and guardrail event is metered inside the sandbox and attributed by root session and task type across an eight-label taxonomy. Analytics can show what the Nemotron-powered agent was doing, rather than collapsing everything into a flat token total.
2. Track the job lifecycle from start to finish. Agentic jobs are declared at the start of a goal arc with a RUNNING marker, stamped as they progress, and closed with an execution status of SUCCESS, FAILED, or CANCELLED. A stale-job janitor closes arcs when an always-on agent is interrupted or halted, which keeps the ledger from stranding open jobs.
3. Attach a business outcome. Revenium's outcome taxonomy turns a finished job into a financial event. A CONVERTED outcome means the business goal was achieved and value was generated, such as a deal closed, a ticket resolved, or a PR merged. A DEFLECTED outcome means the agent avoided a more expensive path through autonomous work. An ESCALATED outcome means a human had to finish the job. CUSTOM covers organization-defined results, while UNSUCCESSFUL captures a completed run that produced no business value. Execution status and outcome type stay separate, because a job can succeed technically and still fail to create value. Revenium tracks both dimensions, including an optional monetary outcomeValue.
4. Compute ROI and unit economics. Once jobs carry outcomes, Revenium's ROI dashboard calculates Value Ratio, which measures business value per dollar of AI compute. It also shows cost per outcome through the conversion and cost-avoidance funnel, with unit economics broken down by agent, model, task type, customer, and product.

NemoClaw runs the agent inside a secure environment. Revenium shows whether that agent is becoming a cost center or a profit center, measured per job, per task type, and per model.

Budget guardrails that hold

In the NemoClaw skill path, budget rules include a warn band, a hard block band, and optional shadow mode for rollout. They can also be created during install from environment variables like REVENIUM_BUDGET_LIMIT and REVENIUM_BUDGET_PERIOD.

A per-turn guardrail directive is injected into every agent turn through an OpenShell before_prompt_build plugin hook. Autonomous-mode installs can also arm a hard halt with channel notifications. That hard halt is available today, while the end-to-end breach-to-halt run on Nemotron is still being validated.

Getting started

Before you install the skill, make sure the environment is ready.

• You have a Linux host ready, whether bare metal, VM, or cloud.
• Docker is installed and running.
• NemoClaw is already installed.
• Your sandbox reports Phase Ready and Inference healthy.
• You have a Revenium API key from app.revenium.ai/connections.

The install takes four steps.

1. Clone the skill into ~/.openclaw/skills/revenium.
2. Export your REVENIUM_* credentials and sandbox name, and optionally a budget limit and period.
3. Run bash scripts/install.sh --nemoclaw.
4. Verify that ✓ ready 💰 revenium shows up in openclaw skills list.

(A note for Mac users: the NemoClaw path is Linux-first today, and macOS support is on the way. macOS users can run the standalone OpenClaw and Docker path now, then move to the NemoClaw path as NemoClaw's macOS support matures. The full runbook, uninstall steps, and troubleshooting live in docs/nemoclaw-setup.md.)

What this gives you

Revenium built this skill because NemoClaw gave agents a secure place to run while leaving the operating economics unresolved.

The skill brings cost, budget, and outcome data into the same sandbox where the agent is doing the work. Operators can see what each job cost, which task type drove the spend, which model handled it, and whether the outcome justified the compute. Budget controls can warn early, block when needed, or run quietly in shadow mode while teams calibrate policy. The sandbox stays intact, and the financial record travels with the agent's work instead of being reconstructed later.

Revenium for OpenClaw inside NemoClaw gives teams a way to run secure agents with a P&L attached.

Start with the NemoClaw skill path. The full implementation and runbook are available on GitHub.