TechOneDigital Request an MCP review

Based on the current stable MCP specification

MCP connects AI to systems. It does not decide what AI should be allowed to do.

MCP standardises how AI applications discover and call business operations across systems. Your organisation still owns authorisation, approvals, evidence and recovery.

Architecture position by , Founder & Technology Consultant.

Protocol fact

What MCP is.

Model Context Protocol is an open protocol for exchanging context and capabilities between AI applications and external systems. In this guide, a business capability means an operation exposed to AI, while a protocol capability is a feature an MCP participant declares it supports. MCP defines participant roles and self-contained request and response messages, including operations for clients to list and invoke tools, read resources and use prompts. An optional discovery method lets a client learn a server's identity and supported protocol capabilities before another request. The specification defines the interface between participants; it does not certify the safety, correctness or business suitability of the operations behind that interface.

Protocol statements on this page use the current stable specification. TechOne positions are explicitly identified as design and operating recommendations.

Protocol boundary

A common protocol is not a complete operating model.

MCP standardises

  • Participant roles and request exchangeMCP defines how hosts coordinate clients, and how clients and servers exchange self-contained request and response messages. The current core has no required handshake or protocol session. Clients include their supported protocol capabilities with each request; servers may expose their identity and supported capabilities through the optional discovery method when a client needs them in advance.
  • Capability discovery and invocationServers can publish tools, resources and prompts through defined protocol primitives so clients do not need a bespoke discovery mechanism for every integration.
  • Structured requests and responsesTool definitions can describe named inputs and outputs with schemas. This creates a common contract surface, although implementers remain responsible for validating and enforcing it.
  • Transport and authorisation patternsThe specification defines supported transports and an OAuth-based authorisation framework for HTTP deployments, including discovery and audience requirements. It does not define each organisation’s business permissions.
  • Extensible protocol capabilitiesThe current specification formalises an extensions model. Tasks are a stable, optional, capability-negotiated extension for work that may outlive a single request. Extensions remain protocol features to assess and govern, not automatic assurance.

The enterprise still decides

  • Decide business authorisationA valid connection or token does not determine whether a caller may submit, post or release an invoice, change production, send a message or disclose a record.
  • Prove that a tool is safe or truthfulNames, descriptions and behavioural annotations help clients reason about a tool, but they are supplied by the server and are not enforcement or independent verification.
  • Create least privilege in the source systemCredential scope, row-level access, field restrictions and segregation of duties must still be implemented in the underlying identity and application layers.
  • Define approval and accountabilityThe protocol does not decide which consequence requires human review, who may approve it, what evidence an approver must see or who owns the resulting decision.
  • Prevent prompt injection or data exfiltrationMCP does not by itself make model inputs trustworthy or stop a capable tool from being misused. Content handling, policy enforcement, isolation and output controls remain architectural responsibilities.
  • Operate the serviceAvailability, monitoring, incident response, contract versioning, reconciliation, rollback and retirement remain part of the organisation’s production operating model.

Illustrative architecture

Keep protocol roles and control roles separate.

A gateway, registry and connector are useful architectural terms, but they are not interchangeable MCP primitives. Naming the role correctly prevents one component from becoming an invisible trust shortcut.

  1. 01AI hostModel, context, user interaction and consent
  2. 02MCP clientOne protocol relationship with one server
  3. 03Optional gatewayRouting, policy hooks, quotas and telemetry
  4. 04Domain MCP serverNamed tools, resources and prompts
  5. 05Source systemExisting API, data and business controls

Identity and policyPrincipal, tenant, scopes, business permission and approval.

Operational evidenceTrace, policy decision, tool version, result class and downstream confirmation.

TechOne reference model. A gateway is optional; source-system authorisation and business rules remain authoritative.

Why this changes integration architecture

Software is becoming discoverable as capabilities, not only navigable as screens.

01 · Protocol fact

A smaller stateless core

The current core does not depend on protocol-level session state between requests. Where continuity is required, state is represented explicitly through mechanisms such as task identifiers or at the application layer, which makes ownership and recovery decisions more visible.

02 · Protocol fact

Extensions become an explicit mechanism

The specification adds a formal extensions framework, allowing capabilities to evolve without placing every feature in the core protocol. Enterprises therefore need an extension acceptance and versioning policy.

03 · Extension fact

Long-running work gains a protocol shape

Tasks are a stable, optional extension for operations that continue beyond a single request. Participants declare and negotiate support through protocol capabilities. Where supported, Tasks improve interoperability but do not decide timeout, cancellation, compensation or accountable ownership.

04 · Protocol and extension fact

Core and enterprise authorisation are distinct layers

The core authorisation specification strengthens issuer validation and client registration guidance. Enterprise-Managed Authorization is a separate stable extension for centrally governed access, not part of the core protocol. Business entitlements and action-level policy still sit outside both.

05 · TechOne position

The differentiator moves from connectivity to control

As protocol support expands, the durable enterprise value is the quality of the capability catalogue: precise contracts, appropriate permissions, reliable refusal, operational evidence and controlled change.

Architecture before implementation

Decisions that should exist before the first production tool.

SDK selection is downstream of these decisions. If they are left implicit, the first connector quietly becomes the identity model, policy model and operating model for everything that follows.

  1. 01

    Is MCP the right integration boundary?

    MCP is especially useful where AI clients need a discoverable, reusable and portable interface to business operations, including when several clients may use the same capability. A stable one-to-one workflow may be clearer and easier to assure as a direct API or conventional integration.

  2. 02

    Where does the trust boundary sit?

    Choose server and gateway boundaries by system ownership, data sensitivity and failure domain. Avoid one universal server that collapses unrelated privileges into a single credential and control plane.

  3. 03

    Whose identity is acting?

    Decide whether an operation uses delegated user identity, a workload identity or a controlled service account. Preserve the initiator, authoriser and executing identity as separate evidence where they differ.

  4. 04

    What data may cross the boundary?

    Define which records, fields and context may enter the model, leave the source system or appear in a result. Apply tenant isolation, purpose limitation, redaction, retention and residency rules before exposing the capability.

  5. 05

    What is the smallest useful capability?

    Prefer named business operations with typed parameters, explicit preconditions and bounded outputs over generic query, script or record-update access.

  6. 06

    Which risk class applies?

    Classify read, reversible write, consequential write and destructive operations. Apply stronger validation, approval and isolation as the consequence increases.

  7. 07

    How can the operation refuse and recover?

    Define validation failures, policy refusals, timeouts, duplicate handling, partial completion, downstream confirmation and rollback or compensation before production release.

  8. 08

    Who owns the contract after launch?

    Assign owners for the tool, source system and control policy. Specify monitoring, evidence retention, version compatibility, deprecation and incident response.

Illustrative architecture scenario, not a customer case.

From request to verified outcome.

Follow one proposed CRM change: a salesperson asks AI to record an agreed next step on an opportunity. The example assumes human review for that field and makes the identity, action and recovery decisions visible.

  1. 01

    Identify the user and opportunity

    Verify the caller and tenant, resolve the exact record and enforce its native CRM permissions.

  2. 02

    Review the proposed change

    Check the tool, arguments and source evidence. Under the example policy, bind approval to the exact record and new value.

  3. 03

    Execute with a command identity

    Use one bounded capability, a stable command ID and the source system’s concurrency controls.

  4. 04

    Confirm or reconcile the result

    Check the source outcome. After a timeout, reconcile the original command before considering a retry; escalate unresolved state.

Before the interface

The boundary can be sound while the data is still unfit for the use case.

If the source, owner, access path, quality or permitted use is unclear, resolve the data decision before designing the MCP capability.

Assess data readiness for one use case

Choose the guide for the decision

Go deeper where the design can fail.

Management, IT leads and readers new to MCP

What Is MCP?

A plain explanation of the protocol: roles, tools, resources and prompts, local and remote servers, and what MCP leaves to your organisation.

Enterprise architects, security and IT leads

MCP Gateway vs LLM Gateway

What an MCP gateway, an LLM gateway and an AI control plane each control, when you need one or both, and what stays with you.

Enterprise architects and platform owners

MCP Gateway Governance

How to use gateway boundaries to govern capability discovery, routing, policy enforcement and evidence across enterprise MCP services.

Security architects, IAM leads and system owners

MCP Security and Authorisation

How identity, credentials, permissions and action-level controls should be designed when MCP capabilities cross enterprise systems.

Engineering leads and service owners

MCP Production Implementation

How to implement, own, observe, version, test and recover MCP services after the first successful connection.

Operational control profiles

The protocol stays the same. The consequence model changes.

These examples show how controls change across operational domains. They are not sector-specific compliance guidance; regulations and internal policy still require a separate assessment.

High-integrity transactional controls

Finance and ERP

Use distinct tools for retrieval, draft preparation, posting and payment. Validate accounting periods, entities, amounts and duplicate keys; preserve segregation of duties; require approval for consequential postings; and confirm the committed state from the source system.

Reversible operations with strict production boundaries

Infrastructure and IT operations

Keep telemetry read-only and separate from configuration changes. Bind tools to environments and approved resources, enforce maintenance windows, capture pre-change state and provide tested rollback. Destructive actions require an independent approval path.

Identity, privacy and outbound consequence controls

CRM, sales and communications

Restrict contact and message data by purpose, record the source evidence, preserve sender identity and distinguish draft creation from sending. Approval should cover the exact recipient, channel and content, with rate limits and a recorded CRM outcome.

Provenance and human accountability controls

Document and regulated workflows

Retain links to authoritative sources, expose uncertainty and exceptions, apply document-level access and retention policy, and separate generation from formal issue. A named reviewer remains accountable for regulated or legally significant output.

Environment and change-management controls

Software and platform engineering

Separate repository reading, branch changes, CI execution and deployment. Keep secrets outside model context, constrain work to approved repositories and environments, require code review for material changes and retain deployment evidence with a recovery path.

Evidence boundary

Protocol claims are cited. Architecture recommendations are labelled.

Because MCP evolves quickly, official sources establish the protocol basis for each technical guide. Optional extensions and implementation recommendations are not presented as core guarantees.

  1. Model Context Protocol specificationPrimary baseline for protocol roles, lifecycle, capabilities, transports and authorisation.
  2. MCP specification release notesOfficial summary of the stateless core, extensions, Tasks and authorisation changes in this baseline.
  3. Tool annotations in the MCP specificationOfficial clarification that annotations are descriptive hints and must not be treated as an enforcement boundary.
  4. Enterprise-managed authorisationOfficial explanation of centrally governed access for enterprise MCP deployments and its relationship to existing identity systems.
  5. MCP Apps authorisation guidanceOfficial guidance on protected resource discovery, OAuth flows and audience-bound access tokens for MCP Apps.

Direct answers

Questions to settle before MCP becomes infrastructure.

Is MCP secure by default?

No. MCP provides protocol mechanisms that can support a secure design, including structured capabilities and an authorisation framework. Security still depends on deployment boundaries, identity, credentials, source-system permissions, validation, approval and operations.

Does MCP replace APIs or integration platforms?

No. An MCP server usually sits above APIs, databases or platform services and presents selected capabilities to an AI client. Conventional APIs, events and workflow engines remain appropriate where discovery by AI clients is not the requirement.

Is OAuth authorisation sufficient for a production MCP service?

No. OAuth can establish and constrain access to a protected resource. It does not decide whether a specific invoice, deployment, message or record change is permitted under business policy. That decision must be enforced at the capability and source-system layers.

Can tool annotations be used as a safety control?

Not on their own. The MCP project describes annotations as hints supplied by a server, not a security boundary. Clients may use them to improve handling, but enforcement must rely on trusted policy and the behaviour of the underlying operation.

When should an MCP action require human approval?

Approval is appropriate when a wrong action would create material financial, legal, operational, privacy or external-communication consequences and automated controls cannot reduce that consequence sufficiently. Approval should show the exact action and parameters, not a generic permission prompt.

Should one MCP server expose an entire business system?

Usually not. Boundaries should follow ownership, sensitivity and failure impact. A smaller set of purpose-built capabilities is easier to permission, test, monitor and retire than a generic interface with broad source-system access.

When is MCP the wrong choice?

It may be unnecessary for a stable one-to-one integration, a tightly deterministic workflow or a use case where no AI client needs capability discovery. Protocol adoption should follow an operating requirement, not precede one.

How does MCP relate to AI Adoption, AI Enablement and AI Agents?

AI Adoption selects valuable priorities, decision owners and working rules. AI Enablement prepares identity, data, platforms and controls. AI Agents use governed capabilities to perform work. MCP can be one interface between those agents and enterprise systems; it is not a substitute for any of the three disciplines.

Use the right commercial entry point

Decide whether MCP belongs in your architecture.

Bring one use case. We will first define the systems, identities, intended capabilities and consequences that belong in the review, then confirm the scope, preparation, timing and price before any work begins.

Request an MCP architecture reviewReview another implementation decision