> This page is for version Latest (default).
> For other versions, use one of these documentation indexes:
> - Latest (default): https://docs.nvidia.com/nemoclaw/latest/llms.txt

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/nemoclaw/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/nemoclaw/_mcp/server.

## v0.0.136

NemoClaw v0.0.136 expands sandbox-attached inference and adds hosted image input and an experimental local profile for Super VL.
It repairs onboarding and DGX Station Express recovery and strengthens provider ownership checks.
Ordinary operation telemetry now uses the fixed TEST receiver.

* Fresh custom hosted endpoint selections bind the validated endpoint, API, and credential source to a provider attached to the selected sandbox.
  Inference status, health checks, switching, and rebuilds use that recorded route.
  Existing shared-route selections remain on their recorded path.
  Related change: [PR #13201](https://github.com/NVIDIA/NemoClaw/pull/13201).
* New Deep Agents Code, OpenClaw, and Pi local selections use sandbox-attached providers for qualified local endpoints.
  Hermes native-local startup, switching, and rebuild refresh supervisor-issued credential references while preserving the selected route and unrelated settings.
  Related changes: [PR #13203](https://github.com/NVIDIA/NemoClaw/pull/13203), [PR #13293](https://github.com/NVIDIA/NemoClaw/pull/13293), [PR #13007](https://github.com/NVIDIA/NemoClaw/pull/13007), and [PR #13292](https://github.com/NVIDIA/NemoClaw/pull/13292).
* For OpenClaw native-local routes, `status`, `connect`, and `start` now send health requests to the provider recorded for the sandbox.
  They no longer probe an unconfigured shared route for those selections.
  Related change: [PR #13447](https://github.com/NVIDIA/NemoClaw/pull/13447).
* An explicit Experimental Nemotron 3.5 Super VL profile serves one 128 GB DGX Spark with native five-token MTP.
  It rejects 64 GB or unknown-memory systems before installation and retains the Spark Qwen default.
  Physical serving and managed-installer smoke tests passed; complete agent onboarding and sustained-load qualification remain pending.
  Related change: [PR #13357](https://github.com/NVIDIA/NemoClaw/pull/13357).
* The NVIDIA Endpoints model picker includes hosted `nvidia/nemotron-3.5-super-vl-120b-a12b` for OpenClaw.
  Interactive and unattended onboarding enable text and image input for that selection unless an explicit valid input override is set.
  Related change: [PR #13371](https://github.com/NVIDIA/NemoClaw/pull/13371).
* Fresh Model Router onboarding waits for the selected OpenClaw gateway to accept connections before requesting its native restart.
  If startup does not settle within the existing bound or the sandbox identity changes, setup stops before restart.
  Related change: [PR #13481](https://github.com/NVIDIA/NemoClaw/pull/13481).
* DGX Station Express keeps automatic model selection under the serving catalog.
  Installation and resume now retain the selected GPU and model, accept the catalog's automatic receipt, verify an owned vLLM server before reuse, and preserve an unchanged sandbox route.
  Related changes: [PR #13305](https://github.com/NVIDIA/NemoClaw/pull/13305) and [PR #13480](https://github.com/NVIDIA/NemoClaw/pull/13480).
* Credential registration inspects existing operator profiles before changing provider state.
  Sandbox creation verifies NemoClaw-owned extra profiles, and messaging setup rejects a fallback to a provider profile it does not own.
  Related changes: [PR #13353](https://github.com/NVIDIA/NemoClaw/pull/13353), [PR #13356](https://github.com/NVIDIA/NemoClaw/pull/13356), and [PR #13351](https://github.com/NVIDIA/NemoClaw/pull/13351).
* OpenClaw rebuild keeps the recorded dashboard origin when it restores a sandbox.
  The `nemoclaw sandbox mcp` usage message now lists the existing `update` action.
  Related changes: [PR #13198](https://github.com/NVIDIA/NemoClaw/pull/13198) and [PR #12384](https://github.com/NVIDIA/NemoClaw/pull/12384).
* The installer and supported CLI operations now attempt anonymous completion telemetry by default through NVIDIA's fixed TEST receiver.
  QA runs can still add a valid test label; `NEMOCLAW_DISABLE_TELEMETRY=1` suppresses collection and delivery, and Production delivery remains disabled.
  Related change: [PR #13494](https://github.com/NVIDIA/NemoClaw/pull/13494).