Komprise wants AI agents to reach every enterprise file store through one MCP server

In Komprise's example, a clinician asks a language model for a specific set of pathology images. The query filters files by KAPPA-enriched context and by that clinician's permissions, then returns the right images ^[12]. A security analyst could use the same interface to find non-compliant files, or to search archived research data by project keyword ^[12]. Both are file-finding jobs. Subramanian said the data behind them can be "millions to billions of files across disparate hybrid storage" ^[5].

On an integration checklist, each vendor's MCP server looks like one more source the agent can reach. Subramanian describes what the agent does with the whole stack differently. "AI is overloaded with multiple tool definitions, resulting in lower accuracy and slower performance rather than a productivity boost," he said ^[4]. The launch coverage does not include a benchmark for that claim, or a price for the product.

The one figure he offered measures a neighbouring problem. He cited a McKinsey study finding that almost 60% of agentic token consumption goes to "response refinement," the loops in which sub-agents process the same data over and over ^[6]. That is repeated processing of data. An agent reading a long list of tool definitions before it acts is a separate cost.

Komprise's product has a part for each problem. For the tool list, it offers one interface to unstructured data wherever it sits ^[1], including storage Komprise does not manage ^[13]. "While different vendors may be adding MCP interfaces to their storage or clouds, Komprise provides a single, unified interface to all the unstructured data in an enterprise across multivendor network-attached storage, object, cloud and application siloes with enriched metadata," Subramanian said ^[14]. For the loops, it sends agents metadata before files and runs noise filters so irrelevant data is not processed ^[10][11]. His complaint about existing connections is that they "haven't tackled the data readiness problem nor high token costs" ^[7].

I think the thing being bought is the index. Permission checks and the shared schema both live in Komprise's layer ^[9]. The person who rolls this out on Monday will be asked on Friday whether that layer's idea of who can open a file matches what the underlying storage says.

Two readings from a team's own agent logs place it on a 2x2. One axis is where the tool list comes from: mostly storage and cloud vendors, or mostly applications the agent takes actions in. The other is whether refinement loops are a large or small share of token spend. Where the list is storage-heavy and loops are heavy, both parts of the product apply and the case for a trial is strongest. If storage vendors dominate the list but loops are light, consolidation is the whole value. The test then is whether the agent picks the right source more often through one interface than through several. An app-heavy list with heavy loops leaves only the noise filters, and only for file queries. With apps dominating and loops light, there is little here to fix. In each case the before-and-after measure is whether the agent returned the right file on the first pass, and how many tokens that took.