Inteligencia Artificial

MCP vs skills/CLI: the authentication isolation that redefines agents

A technical debate reveals that the true value of the Model Context Protocol might simply be externalizing authentication outside the agent's context.

June 21, 2026 · 3 min read

A surreal scene with a silhouetted man approaching a bright arched doorway in a dark room.

TL;DR: Developer Sean Lynch suggests that the main advantage of the Model Context Protocol (MCP) over skills/CLI is isolating authentication outside the agent's context, improving security and simplifying development. The proposal is speculative but technically sound.

What happened?

In a comment on Hacker News, developer Sean Lynch stated that the most valuable capability offered by the Model Context Protocol (MCP) over traditional skills or CLI approaches is isolating the authentication flow outside the agent's context window, and even potentially outside the execution harness entirely. Lynch suggests that the ideal form of MCP could simply be an authentication gateway for the API, and that alone would be a win. The comment was picked up by Simon Willison on his blog, sparking debate in the community. This perspective, though speculative, has reignited discussion about MCP's real role in the AI agent ecosystem.

Why is it important?

Authentication in AI agents has been a critical weak point since the first language model-based assistants. Traditionally, skills or CLI tools handle credentials within the model's context, exposing tokens, increasing the risk of leaks, and forcing frequent reauthorizations. For example, in systems like AutoGPT or LangChain, agents store API tokens in environment variables or even in the prompt, which can lead to leaks if the model regurgitates sensitive information. MCP proposes a standardized protocol where authentication can be managed separately, reducing the attack surface and improving user experience. If MCP is reduced to an authentication gateway, its implementation would be simpler and more secure, but it would also limit its original scope as a universal tool protocol. Historically, the industry has faced similar issues: OAuth 2.0 emerged as a standard for delegating authentication in web applications, and now MCP could do the same for AI agents. However, unlike OAuth, MCP was born with the ambition to also standardize tool invocation and context exchange.

Consequences for the future of agents

This perspective has profound implications:

  • Improved security: By externalizing authentication, agents don't need to store or handle tokens directly, reducing exposure risks. This is especially relevant in enterprise environments where agents interact with APIs containing sensitive data.
  • Simplified development: Developers can focus on agent logic without worrying about credential management. Frameworks like CrewAI or AutoGen could integrate MCP as an authentication layer, accelerating adoption.
  • Standardization: A common authentication gateway would facilitate interoperability between different agents and services. For example, an Anthropic agent and an OpenAI agent could use the same MCP authentication flow, unthinkable with current systems.
  • Possible reduction of MCP's scope: If the protocol is limited to authentication, it loses its ambition to be a complete standard for AI tools. This could lead to fragmentation, where other protocols (like OpenAI's Function Calling) continue to dominate tool invocation.

However, Lynch's vision is speculative. There is no official confirmation that MCP development is heading in that direction. The community is divided: while some see authentication isolation as a pragmatic advance, others believe MCP must offer much more to justify its adoption. It's worth recalling that MCP was introduced by Anthropic in November 2024 as an open protocol for connecting language models with external data sources and tools. To date, it has received contributions from companies like Block and Apollo, but its adoption is still nascent.

What readers should know

This debate reflects the maturity of the AI agent ecosystem. Lynch's proposal is technically sound and addresses a real problem, but it does not represent the official position of Anthropic or MCP's creators. Developers evaluating MCP should consider both its potential security benefits and the limitations if the protocol is reduced to an authentication gateway. MCP's evolution will depend on community contributions and design decisions in the coming months. In parallel, the AI agent market continues to grow: according to Gartner, by 2028, 40% of large enterprises are expected to use AI agents for process automation, and secure credential management will be a key factor. MCP, even as a simple authentication gateway, could become a de facto standard if it simplifies integration with services like Google Workspace, Salesforce, or Slack.

“The true value of MCP might simply be an authentication gateway for the API. That alone would be a win.” — Sean Lynch, via Simon Willison

Keep reading