Workflow guides

Architecture comparison

MCP vs API: choose the interface by the consumer and control boundary

Use an API as the stable system interface for software clients and use MCP when an AI host needs a discoverable, context-rich way to consume selected resources, tools, or prompts. MCP usually sits on top of owned APIs and data access; it does not remove the need for authentication, authorization, validation, rate limits, idempotency, or audit evidence.

Good fit
Choose MCP when multiple compatible AI hosts need to discover and use a bounded capability with descriptions and structured context, while the organization keeps an independent source service underneath.
Pause when
Choose a direct API or internal function when the consumer is deterministic application code, latency and throughput dominate, or exposing model-oriented discovery adds no useful interoperability.

Signals this guide applies

  • The capability should be usable from more than one assistant or agent host.
  • The consumer benefits from resource descriptions and tool schemas at discovery time.
  • The business service already has or needs an interface independent of any model host.

A bounded way to proceed

  1. 01

    Identify the consumer

    List browsers, backend jobs, partner systems, AI hosts, and human operators separately. A single business capability may need a stable API for applications and a narrower MCP adapter for assistant consumption.

  2. 02

    Place the business rule

    Keep validation, permissions, idempotency, and state transitions in the owned service or policy layer. The MCP adapter should translate a model-facing request into that controlled interface, not fork the business logic.

  3. 03

    Compare operational needs

    Assess discovery, compatibility, streaming, batch size, latency, versioning, caching, rate limits, retries, tenant isolation, and traceability. Choose the simplest interface that preserves the required control and service quality.

  4. 04

    Test both boundaries

    Verify the underlying service with deterministic integration tests, then test the MCP resource or tool through a real compatible host. Confirm that errors remain structured and that the adapter cannot expand authority beyond the service contract.

Artifacts an operator can inspect

Consumer and interface map

A diagram ties each consumer to the API, MCP adapter, data source, identity boundary, and owning team, revealing duplicated logic or accidental coupling.

Capability contract

The contract defines business inputs, outputs, errors, side effects, idempotency, service limits, and audit fields independently of the transport used by a client.

Compatibility test matrix

Tests cover direct service calls and supported MCP hosts, including discovery, schemas, authorization failures, partial outages, timeouts, and representative successful cases.

Questions operators ask

Does MCP replace REST or GraphQL APIs?

Usually not. MCP gives AI hosts a standard consumption interface, while REST, GraphQL, RPC, queues, or database access may still power the owned service beneath it. The best layering avoids duplicating business rules in the adapter.

Can the same capability support both MCP and an API?

Yes. Put the capability and policy in a shared service, then expose purpose-built adapters. The API can serve deterministic clients while MCP supplies model-oriented discovery and structured invocation for compatible hosts.

Is MCP only useful for tool calls?

No. MCP can expose resources for context and prompts for reusable interaction patterns as well as tools for actions. Prefer the least powerful primitive that completes the consumer’s job.

AI workflow system

Make the handoff faster without losing the judgment.

Start with one manual handoff. Define the owner, allowed actions, approval pauses, and proof before automation expands.

Verified field result

Automation prepared the evidence. Human judgment still decided.

Routine evidence moved before review. Approval, rejection, and consequential action stayed with a named person.

49/ 50
49 of 50 selected cases produced usable evidence packets for human decision.
Workflow
Marketplace template review
Receipt
#FR-2026-01
Verified
May–June 2026
External writes
0
  1. Map the frictionSee the systems, handoffs, and decision owner.
  2. Test one pathLet routine work move while consequential work pauses.
  3. Keep the proofReview what ran, what stopped, and what needs recovery.