← Back to AI & Technology
Sidy's Intelligence Brief — AI & Technology

MCP: Tool Interoperability Is Not Permission to Act

2026-10-197 min read

The July 28, 2026 MCP specification standardizes tool integration while leaving per-action authorization, result verification and human accountability essential.

MCPAI toolsauthorizationinteroperability

One-Sentence Insight

MCP standardizes how an AI application discovers and calls tools. Technical interoperability does not decide which services to trust or what actions they may perform.

Why the Protocol Matters

A business may connect an assistant to files, databases or management systems. Without shared conventions, every integration needs custom wiring. MCP reduces that friction, while expanding possible actions, data flows and attack surface. Easy installation should never be mistaken for a safe operating boundary.

Explain It Simply

Think of standard electrical plugs. A socket makes connection easy. It does not tell you whether a device is safe, how much energy it will consume or whether it should open your safe. MCP addresses the integration interface; permission and outcome controls remain the host application’s responsibility.

How MCP 2026 Works

The July 28, 2026 revision uses stateless, self-contained requests rather than protocol-level sessions. Servers may expose resources, prompts or schema-described tools. Clients discover capabilities and invoke functions with arguments. The revision also details caching, header-based routing and optional extensions. An application on an older SDK does not automatically adopt the new wire protocol.

Trust and Data Flows

The host application controls model interaction and connections. An MCP client communicates with a server. The server returns definitions and results; some tools can read or change external systems. Retrieved material does not become a higher-priority instruction simply because it arrived through a tool. Trust boundaries exist at each call, permission and result.

Evidence

The official 2026-07-28 specification defines host, client and server roles, JSON-RPC messages and exposed functions. Its Tools section describes parameter schemas and warns that untrusted annotations require caution and human oversight for sensitive actions. The TypeScript SDK migration guide makes clear that 2026 protocol support requires deliberate configuration rather than automatic upgrades.

Economics of Interoperability

Standard interfaces may lower incremental integration cost by avoiding bespoke adapters. Authentication, logging, evaluation, surveillance and maintenance costs remain. A directory of a thousand tools is not necessarily more valuable than ten reliable capabilities. Economic value lies in verified tasks completed, not registry size.

Security Limits

A server may offer dangerous operations or return misleading text. Access controls must consider identity, resource and action. Broken network authorization can expose data; overly broad approval may grant destructive powers. Local transports and remote servers have different credential practices. The protocol does not certify an implementation or a provider's security policy.

What People Miss

Discovering a tool is not granting it permission to act. A 'read-only' description does not prove it cannot modify state. A successful API call is not necessarily a successful business decision. Extensions such as MCP Apps and Tasks are optional, not universal. Compatibility requires testing by protocol revision and capability.

Critical View

MCP primarily solves a common-interface problem. The interface can support excellent or poorly governed architectures. Adding servers without a permission inventory may increase dependencies and risk. A narrowly scoped conventional integration may still outperform a general MCP server. Evaluate rights, costs, observability and actual work performed.

Sidy's Synthesis

My synthesis: do not build a connector collection; build a library of tested capabilities. For every tool, the first record should not ask merely what it can do, but what it can read, change, trigger, share and prove.

I would impose five gates: trusted source, minimum rights, test environment, outcome verification and accountable human owner. A connector that technically works but fails one of those gates remains on hold. That separates an impressive demonstration from a dependable working system.

AI and Future Lens

More autonomous agents may chain tools, move outputs between services and perform long-running work through optional extensions. This multiplies intermediate decisions that users may not see. Monitoring must reconstruct the complete action path, not just the final assistant message.

Conformance testing and enforceable policy could make servers easier to compare. But hostile instructions embedded in external data, privilege drift and vendor dependencies do not disappear when the protocol evolves. Traceability needs to improve as fast as autonomy.

Build From This

Build: an auditable MCP capability register, not an auto-installer. Each record contains owner, authoritative source, protocol revision, transport, operations, permissions, accessible data, cost, reproducible test, observed outcome and approving person.

Pilot five tools needed for a single low-risk workflow, such as retrieving product records or documentation. Test access failures, misleading responses, revocation and unexpected outputs. Keep write operations disabled in the first phase.

Proceed only when tools save time and deliver independently verifiable results. No default broad permissions or automatic installation from unreviewed repositories.

First Steps

Choose one read-only tool. Document what it can access and deliberately test an error case. Do not deploy if the team cannot explain the permission boundary.

Remember This

MCP standardizes connection; it does not certify providers. Tool descriptions are not safety guarantees. Protocol revision and actual permissions require verification.

Primary sources

Facts, figures and quotations should be traceable to the sources below. Sidy's synthesis is labeled as synthesis and does not replace sourced facts.

  1. Model Context Protocol — Official Specification 2026-07-28 (2026-07-28)
  2. Model Context Protocol — Tools, 2026-07-28 (2026-07-28)
  3. Model Context Protocol — Authorization, 2026-07-28 (2026-07-28)
  4. MCP TypeScript SDK v2 — Supporting revision 2026-07-28 (2026-07)