> 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. ## v0.0.107 NemoClaw v0.0.107 adds an Experimental NVIDIA Nemotron 3.5 Lightning profile for one DGX Spark and awaitable Slack readiness for automation. It improves cold-start onboarding, dashboard forwarding, Windows-host Ollama, Hermes operations, sandbox recovery, and CLI inventory output. It also strengthens Hermes Shields state changes and removes sensitive values from more diagnostics. * An explicit-only Experimental managed vLLM profile now runs NVIDIA Nemotron 3.5 Lightning 30B-A3B NVFP4 on one DGX Spark. The profile pins the validated public checkpoint, ARM64 runtime image, parsers, resource settings, and structured serving arguments while the Qwen profile remains the automatic Spark Express default. Managed single-host vLLM setup also stops before credentials, storage, downloads, or container changes when another process owns the serving port. For more information, refer to [Set Up vLLM](/user-guide/openclaw/inference/local-inference/set-up-vllm). Related changes: [PR #8810](https://github.com/NVIDIA/NemoClaw/pull/8810), [PR #8813](https://github.com/NVIDIA/NemoClaw/pull/8813), and [PR #8699](https://github.com/NVIDIA/NemoClaw/pull/8699). * Cold Model Router onboarding now allows a running prefill router up to 10 minutes to download and load its routing model. The startup check stops sooner if the router process exits and terminates a router that does not become healthy before the deadline. For more information, refer to [Set Up Model Router](/user-guide/openclaw/inference/hosted-inference/set-up-model-router). Related change: [PR #8825](https://github.com/NVIDIA/NemoClaw/pull/8825). * OpenClaw and Hermes onboarding now retry the completed OpenShell `sandbox is not ready` forwarding failure after 5 seconds, within the existing three-retry bound. The retry preserves the sandbox and selected port, while authentication, gateway, port-ownership, and unrelated failures remain terminal. For more information, refer to the [OpenClaw Quickstart](/user-guide/openclaw/get-started/quickstart) and [Hermes Quickstart](/user-guide/hermes/get-started/quickstart). Related changes: [PR #8826](https://github.com/NVIDIA/NemoClaw/pull/8826) and [PR #8828](https://github.com/NVIDIA/NemoClaw/pull/8828). * On WSL 2 with Docker Desktop, onboarding now reuses a reachable Windows-host Ollama daemon without applying an unrelated Linux systemd override. The local Linux Ollama path retains its existing override behavior. For more information, refer to [Use Ollama](/user-guide/openclaw/inference/local-inference/set-up-ollama) and [Additional Setup for Windows Machines](/user-guide/openclaw/get-started/additional-setup/windows-preparation). Related change: [PR #8634](https://github.com/NVIDIA/NemoClaw/pull/8634). * `nemoclaw channels status --channel slack --wait` now waits for OpenClaw Slack registration, effective policy coverage, runtime, Socket Mode, and account readiness. Structured output distinguishes ready, waiting, terminal, paused, unsupported, and timeout results without returning tokens or free-text probe errors. For more information, refer to [Set Up Slack](/user-guide/openclaw/manage-sandboxes/messaging-channels/set-up-slack) and the [NemoClaw CLI Commands Reference](/user-guide/openclaw/reference/commands). Related change: [PR #8566](https://github.com/NVIDIA/NemoClaw/pull/8566). * Hermes dashboard pairing and the Hermes gateway now share the durable WhatsApp session path. Sandboxes with credentials in a legacy dashboard path receive cleanup and re-pairing guidance. Hermes skill installs, updates, and removals now direct users to start a new chat session instead of restarting the gateway. For more information, refer to [Set Up WhatsApp](/user-guide/hermes/manage-sandboxes/messaging-channels/set-up-whatsapp) and the [Hermes CLI Commands Reference](/user-guide/hermes/reference/commands). Related changes: [PR #8229](https://github.com/NVIDIA/NemoClaw/pull/8229) and [PR #8535](https://github.com/NVIDIA/NemoClaw/pull/8535). * Current NemoClaw-managed Hermes images on the Docker driver now use the durable runtime provider state mutation contract for Shields transitions. NemoClaw validates that the request and receipt match, holds Hermes gateway startup behind the current fence, and preserves rollback and recovery across controller restarts. An older managed image uses the sealed-plan compatibility path only after NemoClaw proves that the runtime-provider capability is absent. For more information, refer to [Understand Runtime Changes](/user-guide/hermes/manage-sandboxes/configure-sandboxes/understand-runtime-changes), [Troubleshooting](/user-guide/hermes/reference/troubleshooting), and [Trusted Computing Base](/user-guide/hermes/security/trusted-computing-base). Related change: [PR #8658](https://github.com/NVIDIA/NemoClaw/pull/8658). * `nemoclaw rebuild --yes` now reuses a web search credential registered with the OpenShell gateway only when the rebuild hands onboarding a recreate journal for the same sandbox and target intent after deletion, and the live binding matches the sandbox-scoped provider name, provider type, and credential key. Legacy supervisor recovery retries temporary state-backup transport failures after restart, cleans incomplete backups before retrying, and keeps integrity or cleanup failures terminal. Failed OpenShell forward-list queries also remain distinct from an empty list, so cleanup does not stop a forward when it cannot establish ownership. For more information, refer to [Credential Storage](/user-guide/openclaw/security/credential-storage), [Recover and Rebuild Sandboxes](/user-guide/openclaw/manage-sandboxes/operate-sandboxes/recover-and-rebuild-sandboxes), and [Troubleshooting](/user-guide/openclaw/reference/troubleshooting). Related changes: [PR #8774](https://github.com/NVIDIA/NemoClaw/pull/8774), [PR #8787](https://github.com/NVIDIA/NemoClaw/pull/8787), and [PR #8529](https://github.com/NVIDIA/NemoClaw/pull/8529). * `nemoclaw list --json` and global `nemoclaw status --json` now report `openclaw` for a registry entry without an explicit agent, matching the text and sandbox-scoped status surfaces. The commands reference also uses the default `main` OpenClaw agent in agent and session examples instead of `work`, which a clean onboarding does not configure. For more information, refer to the [NemoClaw CLI Commands Reference](/user-guide/openclaw/reference/commands). Related changes: [PR #8710](https://github.com/NVIDIA/NemoClaw/pull/8710) and [PR #8817](https://github.com/NVIDIA/NemoClaw/pull/8817). * A malformed MCP server URL diagnostic no longer echoes the rejected value, which can contain embedded credentials. Managed-image discovery diagnostics now redact every nonempty named token or password value, including values shorter than 10 characters. For more information, refer to [Add an MCP Server](/user-guide/openclaw/manage-sandboxes/mcp-servers/add-an-mcp-server), [Troubleshoot MCP Servers](/user-guide/openclaw/reference/troubleshoot-mcp-servers), and [Security Best Practices](/user-guide/openclaw/security/best-practices). Related changes: [PR #8707](https://github.com/NVIDIA/NemoClaw/pull/8707) and [PR #8744](https://github.com/NVIDIA/NemoClaw/pull/8744).