H
    Hestur
    MCP ConsultingEnterprise Architecture · Governance · Regulated Industries

    MCP as Infrastructure, Not Just Integration.

    For CTOs and engineering leaders who need to think about MCP at the organisational level — not just "can we build an MCP server" but "how do we govern AI tool access across 50 systems and 500 users with the same rigour as our IAM policy."

    The architectural question

    MCP vs direct API integration — when to choose which.

    MCP is not the right answer for every AI integration. It is the right answer when you need standardisation, governance, and multi-agent access to the same systems. Here is the honest comparison.

    DimensionMCPDirect API integration
    DiscoveryTools self-describe via schema — any MCP client discovers capabilities automaticallyRequires manual documentation and custom client code per API
    StandardisationOne protocol — any MCP client works with any MCP serverEvery API is different — REST, GraphQL, SOAP, RPC
    AuditCentral interception point — all calls logged at the serverMust instrument each API client individually
    Access controlRBAC enforced at the protocol layer for all clients uniformlyMust re-implement auth and RBAC in each integration
    AI nativeDesigned for LLM tool calling — structured inputs, typed outputs, error messages designed for model consumptionRequires translation layer between API response format and model expectations
    Multi-agentSame MCP server serves Claude, ChatGPT, Cursor, and custom agents simultaneouslyMust build or maintain separate integrations per AI platform

    Decision framework

    When MCP is the right call — and when it isn't.

    Multiple AI agents need access to the same internal systems

    Use MCP

    One MCP server serves all clients. Without MCP, every agent gets its own bespoke connector.

    Non-technical teams will use AI tools that need live business data

    Use MCP

    MCP exposes controlled, audited access without giving LLMs raw API credentials.

    Compliance requires audit logging of every AI action on business systems

    Use MCP

    MCP servers are the natural interception layer for logging — every tool call passes through.

    Your AI infrastructure will grow to 5+ agents over 12 months

    Use MCP

    Investing in MCP architecture now means each new agent is a configuration change, not a new integration project.

    Simple one-off integration between one tool and one AI agent

    Skip MCP

    Faster to ship, easier to reason about, no protocol overhead.

    Internal tool used only during development or prototyping

    Skip MCP

    Ship fast. MCP is for production infrastructure, not PoC hacks.

    Governance framework

    Enterprise MCP governance — the six pillars.

    MCP at enterprise scale is an infrastructure product, not a project. These six governance areas are what we help large organisations design before the first server is built.

    Tool Cataloguing

    A formal registry of every MCP tool exposed across your organisation: what it does, which systems it accesses, who can call it, what data it touches, and the last security review date. This is the foundation of enterprise MCP governance.

    IAM Integration

    MCP servers authenticate through your existing identity provider (Okta, Azure AD, Auth0). Tool access is tied to existing user roles and group memberships — not a parallel permission system. HR onboarding automatically provisioned the right tool access.

    Data Classification Controls

    Responses from MCP tools are classified by sensitivity (public, internal, confidential, regulated) before reaching the model context window. Confidential data triggers enhanced logging. Regulated data (PII, PHI, financial records) is masked or blocked per policy.

    Audit Trail and Compliance Reporting

    Centralised audit log aggregating tool calls across all MCP servers. SIEM integration (Splunk, Datadog, Azure Sentinel). Monthly compliance reports showing data access patterns, failed authorisations, and anomalies for your GRC team.

    Rate Limiting and Cost Attribution

    Tool call quotas per team, per user, and per agent. Cost attribution dashboard showing which teams are using which AI capabilities and at what volume. Chargeback-ready reporting for enterprise budgeting.

    Version Control and Change Management

    MCP server versions pinned and deployed via your existing CI/CD pipeline. Tool schema changes go through code review and staged rollout. Breaking changes trigger automated tests on all registered AI agent clients before production deployment.

    Regulated industries

    MCP in financial services, healthcare, and legal.

    Financial Services

    Challenge

    AI agents need access to core banking systems, but every API call to production systems is a regulatory event that must be logged, attributed, and auditable.

    MCP Solution

    MCP server layer with read-only tools for account inquiry and fraud detection, write tools requiring dual-approval workflow, full audit trail to SIEM, all data classified and logged per financial services data governance policy.

    Outcome

    AI copilots deployed to 400 relationship managers with full audit compliance on day one.

    Healthcare

    Challenge

    Clinical staff want AI assistance for documentation and research, but PHI access through AI tools creates HIPAA liability without a controlled access layer.

    MCP Solution

    MCP server with patient-record tools scoped to the authenticated clinician's panel only. PHI masked in tool responses before reaching LLM context. Every access logged with clinician ID, patient ID, timestamp, and clinical justification. BAA with all sub-processors.

    Outcome

    AI documentation assistant deployed to 150 physicians with HIPAA-compliant PHI access.

    Legal Services

    Challenge

    AI research tools need access to case management, document repositories, and client data — but attorney-client privilege and matter confidentiality require strict data isolation between matters.

    MCP Solution

    MCP server with matter-scoped tool responses — every query filtered by the requesting attorney's case assignment. Document retrieval logs tied to matter records. Privilege markers propagated from source documents to tool responses.

    Outcome

    Firm-wide AI legal research tool with built-in privilege and confidentiality controls.

    Consulting deliverables

    What you get from an engagement.

    MCP Architecture Document

    Current-state AI integration map, recommended MCP server topology, data flow diagrams, security model, and phased implementation roadmap.

    Tool Registry Design

    Full specification for your organisation's MCP tool catalogue: naming conventions, schema standards, versioning policy, deprecation process.

    Governance Framework

    IAM integration design, data classification matrix for tool responses, audit logging spec, compliance reporting templates for your GRC team.

    Implementation Team Briefing

    Technical deep-dive for your engineering team: MCP protocol internals, TypeScript/Python SDK patterns, testing approach, deployment CI/CD integration.

    Vendor and Platform Evaluation

    Honest assessment of which AI platforms (Claude, ChatGPT, Cursor, custom agents) to prioritise for MCP integration based on your current stack and roadmap.

    Build-or-Buy Analysis

    For each proposed MCP server: build internally vs. use a third-party connector vs. engage us for development. Priority and sequencing recommendations.

    Ready to design your MCP infrastructure?

    90-minute architecture session with your engineering and security leadership. We map your AI systems, design the MCP layer, and produce a prioritised implementation plan. Fixed-price engagement.