Expert AI Layer and MCP

MCP, the Model Context Protocol, provides a standard way for an AI application to connect to tools and knowledge. An Expert AI Layer provides the maintained professional context that the AI application should use.

MCP can be the interface between an AI client and expert knowledge. It does not replace the knowledge model, governance, search strategy, orchestration, or expert ownership behind that connection.

What is MCP?

MCP is a protocol for connecting AI hosts and clients with servers that expose resources, tools, and prompts. An AI application can use the connection to discover available capabilities and request information or operations.

The exact implementation depends on the server and client, but the architectural distinction is important: MCP defines a way to connect, not the professional meaning of the knowledge being connected.

What MCP gives an AI system

a standard connection to external capabilities;
access to selected resources and knowledge;
tools for search, reading, or controlled operations;
prompts that help an AI client use the connection;
a boundary between an AI application and a knowledge or tool server.

This makes MCP useful for connecting an AI client to a maintained Expert AI Layer without embedding all knowledge in the model or one application.

What MCP does not define

MCP does not decide which content is accepted knowledge, who owns it, whether a rule is current, how exceptions work, or when a human must approve a decision.

It also does not replace:

the expert knowledge model;
status and version governance;
semantic and exact search strategy;
permissions and tenant isolation;
agent orchestration;
the AI model’s reasoning;
human responsibility for important decisions.

How MCP and Expert AI Layer fit together

Expert and company knowledge
        ↓
Expert AI Layer: structure, status, scope, relationships
        ↓
Noda MCP server: controlled resources and tools
        ↓
MCP client and AI host
        ↓
Answer, recommendation, or action

The server should expose the right context for the task rather than returning an unfiltered archive. The Expert AI Layer determines what the content means; MCP makes selected content available.

MCP as an interface to knowledge

A knowledge server can expose search and read operations so an AI client can request relevant material. The response should identify the article, category, status, and source context where appropriate.

Search should distinguish accepted knowledge from drafts, rejected material, and superseded guidance. It should also respect the user’s category and operation permissions on the server.

Search for accepted knowledge

The AI client may ask a broad question, but the server should apply deterministic filters, category scope, tenant scope, and access rules before returning results. Semantic similarity alone is not enough. A similar but outdated or unauthorized article is not a safe answer.

Why an MCP tool should not return everything

A universal “give me all knowledge” tool creates huge context, increases contradictions, and makes it difficult to know why a result was selected. Narrow tools produce clearer requests and more predictable permissions.

A search tool should return relevant candidates and metadata. A read tool should return the requested accepted article or clearly report that no applicable accepted knowledge was found.

Resources, tools, and prompts

Resources

Resources can expose stable, human-readable knowledge or reference material. They are useful when the AI needs to inspect content without performing an operation.

Tools

Tools perform a defined request such as search, read, suggest, or a controlled write. Tool names and inputs should reflect narrow jobs.

Prompts

Prompts can guide an AI client to search before answering, report gaps, cite knowledge, or request human review. They do not replace the underlying expert content.

Example interaction

The user asks an AI client how to handle a particular customer case.
The AI searches the relevant Noda category through MCP.
The server filters by tenant, permissions, status, and scope.
The AI reads the accepted method, rule, and linked exception.
The AI asks for a missing fact if the method requires it.
The AI prepares an answer or recommendation.
If action authority is missing, the AI reports the required human approval.

What happens when knowledge is absent?

The server should not return a confident substitute when no accepted knowledge applies. It should report no result, identify the missing category or condition, and allow the AI to ask a focused question or escalate.

Missing knowledge can become a suggestion for expert review. It should not automatically become an accepted article.

Access and security

Access must be enforced on the server, not only described in a prompt. The server should validate the tenant, access key, category, operation, and requested article before returning content.

Read access and write access should be separate. Access to a knowledge category does not automatically grant permission to suggest, approve, delete, or change it.

Read, suggest, and govern

Read

Search and read accepted knowledge for an AI task.

Suggest

Propose a new article, correction, relationship, or exception for human review.

Govern

Approve, reject, update, restore, or delete knowledge under explicit authorization. AI should not approve its own drafts.

One server or several?

One Noda MCP server can serve several categories and AI clients when tenant scope, permissions, and operations are properly separated. Separate servers or gateways may be useful for different security domains, environments, or organizational boundaries.

A gateway can provide a controlled entry point while routing requests to specialized servers. The choice depends on access, scale, audit, and operational requirements.

Why categories matter

Categories define the boundary of expert context. They help a server restrict search and read operations, keep unrelated knowledge apart, assign domain owners, and measure how a particular AI workflow performs.

A category is more useful than one universal knowledge stream because it connects content with scope and responsibility.

Designing MCP tools for an Expert AI Layer

Keep tasks narrow. Separate search, read, suggest, and governance operations.
Return status and identifiers. The AI should know what it read and whether it is accepted.
Report no result clearly. Absence is meaningful and should not be hidden.
Filter on the server. Never rely on the prompt to enforce permissions.
Separate draft and accepted. Drafts should not appear as ordinary authoritative results.
Support audit. Record important reads, suggestions, approvals, and changes where appropriate.

MCP does not replace orchestration or a data model

MCP connects the client and server. Orchestration decides when to search, how to combine results, when to ask a question, and when to call a tool. The data model defines articles, categories, relationships, statuses, ownership, and versions.

A strong system needs all three: a meaningful knowledge model, a reliable connection, and an AI workflow that uses them responsibly.

How Noda supports this architecture

Sekura Noda can act as a maintained Expert AI Layer and expose controlled knowledge operations through MCP. Noda’s role is not only to transport text. It helps preserve articles, categories, statuses, relationships, access keys, and tenant-scoped operations.

Common mistakes

Calling MCP a knowledge base

MCP is a connection protocol. The knowledge model and governance remain separate concerns.

Creating one universal tool

Broad tools make permissions, intent, auditing, and context selection harder.

Allowing AI to approve its own drafts

Suggestions require an accountable owner before becoming accepted knowledge.

Returning rejected material as ordinary results

Status must be visible and respected by search and read operations.

Failing to filter on the server

Prompts are not an access-control system.

Considering connection the finished product

A working MCP connection is only the interface. Expert knowledge still needs structure, review, ownership, and development.

Frequently asked questions

Is MCP required for an Expert AI Layer?

No. MCP is one useful connection method. An Expert AI Layer can use other interfaces, but MCP provides a standard way to expose knowledge and tools.

Can an MCP server store knowledge itself?

It can, but storage is not the same as expert governance. The system still needs status, ownership, scope, relationships, and review.

What should be a resource and what should be a tool?

Use resources for stable content and references. Use tools for defined operations such as search, read, suggest, or governed changes.

Can one MCP server serve several models?

Yes, when access, categories, operations, and tenant scope are enforced consistently.

Should articles include a protocol version?

Technical pages should identify the version and verify details against the current official specification before publication.

Conclusion

MCP connects AI applications to knowledge and tools. The Expert AI Layer supplies the professional context, rules, methods, exceptions, status, and boundaries that make the connection useful.

Use MCP as a controlled interface to expert knowledge, not as a replacement for the knowledge model or human governance.