Release Notes

August 4, 2026

v0.0.102

NemoClaw v0.0.102 adds authenticated attachment of operator-managed llama.cpp servers and an Experimental managed vLLM profile for two DGX Spark systems. It also improves DGX Station and Windows installation, gateway and sandbox recovery, Shields transactions, inference reliability, multi-port uninstall, and Hermes and LangChain Deep Agents Code workflows. It reduces the managed single-node DGX Spark Qwen profile’s default memory demand during long-context agent workflows.

  • The nvidia/Qwen3.6-35B-A3B-NVFP4 profile no longer enables multi-token prediction (MTP) speculative decoding by default. It retains async scheduling and all other registered serve arguments. Operators can apply lower context, concurrency, and batch limits for bounded long-context investigation, but the override does not guarantee protection against a host freeze. For more information, refer to Set Up vLLM.
  • Onboarding can now attach an authenticated, operator-managed llama.cpp server on loopback port 8081 as the llama-cpp-local provider. NemoClaw requires consistent native fingerprint evidence, bounds probe responses, and retains the generic compatible-endpoint path when the server cannot be identified as llama.cpp. For more information, refer to the NemoClaw CLI Commands Reference.
  • Experimental Express installation can now discover two qualified DGX Spark systems and select their managed vLLM profile from the declarative serving catalog. DGX Station paths also qualify the May 2026 GB300WS factory image, retain the qualified driver transaction, accept mode-bound Express resume state, and recover host-global dual-Station runtime ownership during later onboarding. For more information, refer to Set Up vLLM on Two DGX Sparks, Set Up vLLM on Two DGX Stations, and Additional Setup for DGX Station.
  • Windows Subsystem for Linux (WSL) onboarding now validates Windows-host Ollama from Docker Desktop’s network context and reuses a running daemon on mirrored networking. When WSL has no local ollama executable, model pulls use the daemon’s HTTP API instead. For more information, refer to Additional Setup for Windows Machines and Use Ollama.
  • Reinstallation now reuses a healthy installer-managed CLI when its selected source and build identity still match. Linux installation rejects incompatible OpenShell gateway versions before onboarding, adopts only a compatible package-managed gateway, and permits the bounded package-service-to-standalone recovery transition. Multi-sandbox onboarding also records the dashboard port selected after a collision or final recovery. For more information, refer to Gateway Lifecycle Authority and the NemoClaw CLI Commands Reference.
  • Starting a stopped sandbox now restores managed startup state before final gateway and host-forward readiness checks. The gateway watchdog recovers every classified not-serving state after the gateway has served, and managed recovery applies NEMOCLAW_GATEWAY_RECOVERY_WAIT_SECONDS consistently. Docker sandbox creation also retains rollback authority through late readiness, GPU, and local-inference checks. For more information, refer to Run Sandboxes and Recover and Rebuild Sandboxes.
  • Shields deadline recovery now blocks new mutations, waits for the verified owner to release its lock, and enters durable containment when bounded recovery cannot establish authority. Relock can repair a narrowly validated permission-only configuration-hash drift, while corrupt transition locks fail promptly with recovery guidance. For more information, refer to Understand Runtime Changes and Create and Restore Snapshots.
  • policy restore now accepts --yes, -y, and --force, previews restored egress or stale-record cleanup, and revalidates the target before mutation. Explicit sandbox destruction can continue after pre-delete Shields hardening fails while preserving the auto-restore authority if deletion is not confirmed. The new nemoclaw uninstall --all-gateway-ports mode processes every discovered gateway-port environment independently and preserves shared resources when cleanup remains incomplete. For more information, refer to Network Policies, Uninstall NemoClaw, and the NemoClaw CLI Commands Reference.
  • OpenClaw now retries one classified transient remote Model Context Protocol (MCP) startup with a fresh transport and does not retain a degraded tool catalog as the stable session catalog. Hosted inference probes use a bounded reply budget accepted by endpoints with a higher minimum, and onboarding preserves the validated reasoning-capability value through resume and sandbox creation. For more information, refer to Troubleshoot MCP Servers and Configure Model Capabilities.
  • Hermes WhatsApp pairing and media traffic now use the injected OpenShell proxy. Hermes sessions delete invokes the native Hermes command with native identifier validation. Deep Agents Code applies configured corporate certificate authority (CA) trust during local base-image dependency fetches and in the final sandbox image. For more information, refer to Set Up WhatsApp, the Hermes NemoClaw CLI Commands Reference, and Configure Corporate CA Trust.
  • Managed OpenClaw, Hermes, Deep Agents Code, and MCP discovery images now carry reviewed dependency updates for the release advisory set. The changes update locked runtime graphs, remediate private npm dependencies during image assembly, and preserve fail-closed identity and integrity checks. For more information, refer to the OpenClaw 2026.7.1 Dependency Review and Hermes 0.19.0 Dependency Review.
  • Affected CLI paths now report invalid enumerated values, unresolved sandbox base images, and changed gateway authority without raw Node.js stack traces. Captured output from non-JSON OpenClaw agent commands is inspected for recognized embedded-fallback markers. When a marker is present, NemoClaw suppresses the captured transport output, exits with status 1, and prints the documented recovery commands. For more information, refer to the NemoClaw CLI Commands Reference and Troubleshooting.
  • Documentation now gives OpenClaw, Hermes, and Deep Agents Code the same one-command interactive installation start and describes the current product capabilities and boundaries. For more information, refer to the OpenClaw Quickstart, Hermes Quickstart, Quickstart with LangChain Deep Agents Code, and Platform Support.