代理基礎設施
NemoClaw 0.0.113 Strengthens Local Inference Validation, Can Reissue MCP Credentials After Sandbox Rebuilds
NVIDIA has updated its open-source agent execution stack, focusing on state drift across vLLM, llama.cpp, Ollama, and the model router. The new release also tightens sandbox lifecycle management, image provenance, and failure recovery, though it remains a patch release in a rapidly evolving project.
NVIDIA updated the [NemoClaw release notes](https://docs.nvidia.com/nemoclaw/user-guide/deepagents/release-notes/2026/8/20) on August 20, making state synchronization issues among local models, sandboxes, and agents a primary target for fixes. NemoClaw is an open-source reference stack built on OpenShell. It can run OpenClaw, Hermes, and LangChain Deep Agents Code inside sandboxes governed by network and file policies, while consolidating inference routing, MCP integration, snapshots, and lifecycle operations in a single CLI.
The managed vLLM workflow in 0.0.113 no longer reuses an existing process merely because a service is listening on the port. It now checks the serving profile, model, inference mode, and port configuration. Interrupted installations can also resume while preserving the selected provider. Ollama validation now confirms the actual model through the sandbox endpoint, Portable mode can use Podman, and Windows/WSL detection has been bounded to prevent credential rejections from being misclassified as routing failures. The llama.cpp bridge adds validation and request-size limits, and can connect to servers that do not report model aliases through `/props`.
On the lifecycle side, the new release creates the Portable network before writing the host alias and can rebuild the gateway if the Docker network disappears. After policies are bound, it republishes MCP credentials and the provider reservation. Interactive `connect` sessions release their lock before waiting for the shell, preventing other commands for the same sandbox from being blocked for extended periods. Production image requests now also require immutable semver tags and complete source commit attestations. The implementation and installation instructions can be reviewed in the [NemoClaw GitHub repository](https://github.com/NVIDIA/NemoClaw).
These changes matter for self-hosted agents because they introduce more cross-validation between the declared configuration and the models, networks, and credentials actually in use. However, the release is focused almost entirely on reliability and software supply chain controls, with no quantified improvements in latency, security, or failure rates. Given the rapid release cadence, production environments should continue to pin tags, validate snapshot restoration and policy migration first, and only then upgrade existing sandboxes.