Xinnor has released xiNAS 3.13, the new production version of its NFS over RDMA storage platform, alongside two technology previews that map its roadmap: AI-assisted storage management through MCP in xiNAS 3.14, and live storage-condition exposure for pNFS coordination software in xiNAS 3.15.

The announcement, dated Sept. 28, 2026, arrives months after xiNAS debuted on Supermicro storage servers with AMD EPYC processors and PCIe Gen5 NVMe drives. A single 12-drive system demonstrated up to 74.5 GB/s read and 39.5 GB/s write, while a two-server configuration reached 117 GB/s read and 79.6 GB/s write — figures that anchor the platform's positioning for data-intensive AI and HPC workloads.

One control plane

The core change in 3.13 is a single control plane spanning storage provisioning, NFS configuration, health inspection and configuration history. Whether an administrator works from the terminal interface, CLI, REST API or an MCP client, xiNAS follows the same workflow: it builds an end-to-end plan for a given task showing what will change, checks permissions, applies the approved operation and records it in the audit log.

Operations that can alter or destroy data require the appropriate role and explicit confirmation. The intent is to give administrators one consistent path for managing RAID, filesystems, NFS and networking, rather than separate management interfaces for each layer.

AI without shell access

The xiNAS 3.14 technology preview extends that control plane to MCP-compatible AI assistants and automation tools. Administrators can direct an AI client to inspect system health, investigate unavailable shares, review changes or prepare configuration operations.

Xinnor draws a hard line at autonomy: the AI client does not receive shell access. The platform maps requests to defined xiNAS operations, which remain subject to the same permissions, planning, confirmation and auditing rules as any other management interface.

Diagnostics also get a stricter treatment of absence. Results record which checks succeeded, failed, timed out or could not be performed, so missing information is never treated as evidence of a healthy system — a meaningful distinction in AI-driven operations where an assistant may otherwise read silence as success.

"AI can make storage administration much easier, but it should not bypass the controls that make infrastructure predictable and safe," said Sergei Platonov, Vice President of Strategy at Xinnor. "In the 3.14 preview, an assistant can inspect the system, explain what it finds and prepare an operation, while permissions, the execution plan and final confirmation remain under xiNAS control."

Platonov will present the architecture and safety model behind the MCP preview at the SNIA Developer Conference 2026 on September 30.

Placement aware of storage condition

The 3.15 preview addresses a different problem in scale-out pNFS environments: how a coordination layer can account for the actual condition of storage servers when placing new data.

In the preview, xiNAS publishes the condition of the RAID, filesystem and NFS layers behind each share. A connector can feed this information to pNFS Lattice, an open-source metadata and coordination layer for pNFS. The result is a basis for evaluating placement policies that could steer new data away from servers that are rebuilding, degraded or unavailable.

Xinnor frames the 3.15 work explicitly as an early technology preview for technical evaluation and development — it is not part of the current production xiNAS feature set.

Availability

xiNAS 3.13 is available now as the current stable release. Source branches for 3.14 and 3.15 are available separately for technical evaluation. Xinnor is also nominated in the 2026 HPCwire Readers' Choice Awards in the Best HPC/AI Storage Product or Technology and Top 5 New Products or Technologies to Watch categories.

Whether the MCP-based management model becomes a differentiator will depend on how quickly enterprise and HPC shops trust AI assistants with infrastructure operations — but Xinnor's decision to keep permissions, execution planning and final confirmation inside xiNAS, rather than in the AI client, sets the safety terms other storage vendors will likely be measured against.