Sandboxes and scanners

Pipelock vs NemoClaw

Content scanning next to a sandbox reference stack. NemoClaw decides where an agent may connect. Pipelock inspects what it sends.

At a glance

Pipelock source NemoClaw
Job Agent firewall. Mediates HTTP, WebSocket, and MCP traffic routed through it, scans it for secret leaks, prompt injection, SSRF, and tool poisoning, and can emit signed action receipts for mediated decisions when a signing key is configured. Reference stack for running supported agents inside NVIDIA OpenShell sandboxes: guided onboarding, managed inference, network policy with operator approval, integrations, snapshots, and lifecycle.
Enforcement point Network path, outside the agent process The sandbox around the agent process and its egress allowlist
Source Open source, Apache-2.0 core; Enterprise under ELv2 Open source, Apache-2.0 (NVIDIA); described by NVIDIA as an alpha project
Pricing shape Free core; paid Pro and Enterprise tiers Free
Runs as Single Go binary, self-hosted; container and Helm CLI on a host with a supported container runtime; agents run inside OpenShell sandboxes
Pick NemoClaw

You run OpenClaw, Hermes, or LangChain Deep Agents and want a hardened sandbox with an egress allowlist and an approval flow for new destinations.

Pick Pipelock

You need to inspect the content of traffic routed through Pipelock across MCP, HTTP, and WebSocket, and keep signed records where receipt signing is configured.

Run both

NemoClaw draws the execution boundary. Pipelock inspects traffic routed through it. Running both covers both boundaries.

Want the runtime boundary, not just another checklist?

The short version

Pipelock is an open-source agent firewall. It scans HTTP, MCP, and WebSocket traffic routed through it for credential leaks, prompt injection, SSRF, and tool poisoning. With a signing key configured, it can emit signed action receipts for mediated decisions. Single binary, works with any agent routed through it.

NemoClaw is NVIDIA’s open-source reference stack for running supported AI agents more safely inside OpenShell sandboxes. It provides guided onboarding, managed inference, network policy, managed integrations, snapshots, and lifecycle operations through a CLI. The README lists OpenClaw as the default agent and Hermes and LangChain Deep Agents Code as supported, and calls the project alpha with best-effort maintenance.

They operate at different layers. NemoClaw controls the execution environment and which destinations the agent may reach. Pipelock inspects traffic routed through it on the way to those destinations.

Feature comparison

FeaturePipelockNemoClaw
ArchitectureNetwork proxy, single binarySandbox orchestrator on top of OpenShell
Agent supportAny agent routed through the proxyOpenClaw, Hermes, LangChain Deep Agents Code
Credential scanning (DLP)65 patterns, encoding-aware, on mediated trafficNot described in public materials
Prompt injection detectionDeterministic patterns with multi-pass normalization on mediated trafficNot described in public materials
SSRF protectionMediated private-IP, metadata, and DNS rebinding checksNot described in public materials
MCP scanningMediated traffic, both directions: tool descriptions, arguments, responsesNot described in public materials
WebSocket scanningMediated frames, both directionsNot described in public materials
Network policyDomain blocklist and allowlist plus content scanning on routed trafficDeny by default; baseline endpoint rules scoped to a binary with method and path rules; Restricted, Balanced, Open, and Personal tiers of presets; operator approval flow for unlisted hosts
Broad-egress modeNo broad egress policy; routed traffic still passes Pipelock’s scannersPersonal tier: any host on ports 80 and 443, which the docs say is not inspected for protocol or payload and does not prompt
Process sandboxLandlock, seccomp, and network namespaces (process-level)OpenShell sandbox (container-grade)
Managed inferenceNot in scopeYes
Snapshots and lifecycleNot in scopeYes
Signed receiptsAvailable with a signing key; Ed25519, verifiable offlineNot described in public materials
LicenseApache-2.0 core; Enterprise under ELv2Apache-2.0

Where NemoClaw is stronger

Container-grade isolation. The agent runs inside an OpenShell sandbox. Pipelock’s own sandbox is process-level, and Pipelock doesn’t manage a container lifecycle.

Operator approval for egress. The network-policy reference describes a deny-by-default baseline, rules that name the endpoint, the requesting binary, and the allowed methods and paths, and a TUI approval flow: an unlisted host is blocked and logged, shown with host, port, and binary, and added to the running policy if approved. That’s a useful gate for interactive agents, and it is finer than a hostname allowlist.

Managed inference and onboarding. Model routing, snapshots, and a guided setup for the supported agents. Pipelock doesn’t route model calls or manage agents.

Where Pipelock is stronger

Content inspection. An allowlist decides which hosts the agent may reach. It doesn’t read what the agent sends to an allowed host. NVIDIA’s own reference says so for the Personal tier: the rule “does not inspect the application protocol or payload,” and “an agent can send workspace data or sandbox-visible credentials to an arbitrary reachable service on either port without an operator approval prompt.” When the sandbox routes traffic through Pipelock, Pipelock scans the request body for encoded credentials, the response for injection, and the tool description for poisoning.

Transport coverage. Pipelock mediates HTTP forward proxy traffic, HTTPS via CONNECT, WebSocket frames, and MCP over stdio and HTTP, with the same scanning engine on each.

Framework independence. Pipelock works with any agent whose traffic goes through it, including agents NemoClaw doesn’t support.

Evidence. With a signing key configured, Pipelock can emit signed receipts for mediated decisions. The shipped verifier checks them offline.

The enforcement gap

NemoClaw:  Agent -> [sandbox + egress allowlist] -> allowed hosts only
Pipelock:  Agent -> [content-scanning proxy]     -> scanned routed traffic
Both:      Agent -> [sandbox] -> [Pipelock]       -> scanned routed traffic to allowed hosts

The third line puts the controls in series. NemoClaw provides the execution boundary. Pipelock provides the content boundary for routed traffic. Different failure modes, one stack.

A note on what this page could verify

The NemoClaw column is drawn from the README, the docs index, and the Network Policies reference page, each dated above. The security best-practice pages were not read for this revision, so nothing here rests on them.

Further reading

Sources checked

Third-party descriptions on this page come from the public materials below, read on the dates shown. Features and pricing change; check the current documentation before you decide.

  • NVIDIA/NemoClaw README checked 2026-09-01 · supported agents listed as OpenClaw (default), Hermes, and LangChain Deep Agents Code; alpha status stated in the README
  • NemoClaw documentation checked 2026-09-01 · docs index
  • NemoClaw docs: Network Policies reference checked 2026-09-01 · deny-by-default; baseline policy allows named endpoints per binary with method and path rules; Restricted, Balanced, Open, and Personal tiers; Personal is a hostless port 80 and 443 rule that the page says does not inspect protocol or payload and does not prompt; external inference goes through the OpenShell gateway; operator approval flow in the TUI

Third-party product names and marks belong to their owners. PipeLab is not affiliated with, sponsored by, or endorsed by the makers of any product compared on this page. Descriptions of other products come from their own public materials on the dates listed above and reflect PipeLab's reading of them. If something here is wrong or out of date, tell us and it will be corrected.

Frequently asked questions

What's the difference between Pipelock and NemoClaw?
NemoClaw is NVIDIA’s reference stack for running supported agents inside OpenShell sandboxes, with network policy, managed inference, and lifecycle tooling. Pipelock is a network proxy that scans the content of HTTP, MCP, and WebSocket traffic routed through it. NemoClaw controls where the agent can connect. Pipelock inspects what the agent sends through the paths routed to Pipelock.
Does NemoClaw scan for credential leaks or prompt injection?
Its network-policy reference describes a deny-by-default policy, baseline endpoint rules scoped to a binary with HTTP method and path rules, tiers of presets, and an operator approval flow. It also states that the Personal tier rule does not inspect the application protocol or payload. Content scanning of request or response bodies, DLP pattern matching, or injection detection is not described in the public materials read for this page.
Can I use Pipelock and NemoClaw together?
Yes. NemoClaw provides the sandbox and allowlist. Pipelock can run as the proxy the sandboxed agent talks through and scan that content. The deployment can use both containment and content inspection.

Want the runtime boundary, not just another checklist?

See all comparisons →