MCP
Model Context Protocol layer.An open standard for giving an agent the right context at the right time. In CORTX, MCP is the layer that turns captured knowledge into usable agent runtime.
An open protocol with an opinionated implementation.
Model Context Protocol is an open standard for connecting large language models to data, tools, and external systems in a consistent way. It defines how a model asks for context, how a context provider responds, and how the connection is structured.
CORTX implements MCP at the heart of every deployment. The standard handles the protocol; CORTX provides the layered, persistent, customer-owned context that fills it. The protocol is generic. The contents are specific to your business.
Four layers of context.
The four layers compile into the agent's runtime context for every action.
What happens when the agent acts.
When the agent receives a task, it queries the MCP layer. The query specifies which workflow is active, which organization the task belongs to, and which operator is assigned. The MCP layer assembles a runtime context from the four layers — System for the workflow primitives, Org for the configuration, Employee for the operator's preferences, Vertical for the cross-organization patterns — and returns it.
The runtime context is what the agent uses to decide what to do next. It is not a prompt template. It is structured data — workflows, primitives, vocabulary, exception rules — that the agent reads as part of every loop iteration.
Updates to any MCP layer take effect on the next loop iteration. There is no deployment, no reload, no downtime.
Memory as files, not as a database.
The most consequential decision in an agent's architecture is where its memory lives. Most agent systems put memory in a vector database, an opaque embedding, or a hosted state service. Each of those choices makes the system harder to inspect, harder to audit, and harder to control.
CORTX puts memory in files. MCP files are versioned, diffable, inspectable, and rollback-able. They live on the customer's machine. They can be read by a human in a text editor.
The trade-off is real. Files are slower than vectors for some kinds of retrieval. But files are auditable, and vectors are not, and in operational AI auditability is not optional.
Between the vault and the agent.
MCP sits between the static vault — where the four layers live as files on disk — and the agent runtime — where decisions get made. The vault is the long memory. The agent is the actor. MCP is the protocol that connects them.
Where MCP shows up.
Captures the raw material that becomes System MCPs.
Read Flow One Place. Every Case.Runs against the assembled MCP context for every operator.
Read Healthcare Clinics, labs, insurance claims.A new set of MCPs built on the same protocol.
Read Professional Services Accounting, legal, consulting.A new set of MCPs built on the same protocol.
Read Distribution Importers & wholesalers.A new set of MCPs built on the same protocol.
Read Logistics Freight, warehousing, fulfillment.A new set of MCPs built on the same protocol.
Read