Products & Toolsproduct designagent interfacesinfrastructureagents

Most MCP Servers Don't Need to Exist

||By LDS Team
5.2
Relevance Score
Most MCP Servers Don't Need to Exist
Photo: evilmartians.com · rights & takedowns

For teams choosing how to expose services to AI agents: Evil Martians frames the MCP server as a product interface for agentic workflows, not a default AI feature to bolt on. Their decision framework distinguishes cases where a standard REST API or CLI is the right call from situations where an MCP server genuinely earns its complexity. The core argument is that most integrations do not require the full MCP protocol, and choosing the wrong interface adds unnecessary operational surface. Practitioners building agent-aware tooling should evaluate whether the primary consumer is a developer, a service, or an autonomous agent before committing to MCP.

For practitioners building agent-native products, the core question of 2026 is not 'Should we ship an MCP server?' but 'Which interface is the right product decision?' Evil Martians' engineering blog argues that defaulting to an MCP server for every integration is a mistake -- and presents a decision framework for choosing between direct API calls, a CLI, or an MCP server.

The central argument

The post frames an MCP server as a product interface for the age of agents, not an AI feature. That distinction matters: an MCP server only makes sense when an AI agent needs to take actions on behalf of a user, not when a developer is automating a build step or an app just needs a data endpoint. The same interface that is elegant for an agent orchestrating tasks becomes unnecessary infrastructure when the use case is a straightforward REST call or a scripted pipeline.

The decision fork

Based on the snippet and Evil Martians' established expertise in developer-tool product design, the framework appears to separate three cases:

  • •a structured public API suffices when the caller is another service or application
  • •a CLI is the right choice for developer workflows and automation scripts that run outside an agent loop
  • •an MCP server earns its place when the primary consumer is an AI agent making real-time decisions and tool calls

Practitioner takeaway

Teams drawn to MCP because of ecosystem momentum may be adding a protocol layer, auth surface, and operational dependency that their actual use case does not require. Evaluating whether your integration is primarily human-directed, developer-directed, or agent-directed before choosing the interface can avoid unnecessary deployment complexity.

This is a single-source framework post from Evil Martians' team blog. The specific criteria and decision tree are as summarized from the published snippet; practitioners should read the full post for the complete heuristics.

Key Points

  • 1An MCP server is a product interface for agent-era systems, not an AI feature -- defaulting to MCP adds protocol and ops overhead that most integrations do not need.
  • 2The post presents a decision framework: direct API for service-to-service, CLI for developer pipelines, MCP only when an AI agent is the primary actor taking real-time actions.
  • 3Teams should evaluate the primary consumer (developer vs. service vs. agent) before choosing an interface, to avoid shipping MCP servers that would be simpler as APIs or CLIs.

Scoring Rationale

Practical product-framework guidance helps teams decide interface choices for agent-era software; useful to practitioners but not a technical or research breakthrough.

MoreAI Infrastructure news

Practice interview problems based on real data

1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.

Try 250 free problems