Glonce

A blog post argues MCP has already outlived its usefulness

agentsThe Glonce editorsHacker News🔥59Glonce rating59Impact45Novelty35Relevance88Surprise48Players60Credibility65

A blog post titled "Why MCP Was Always a Bad Idea" argues that the Model Context Protocol, the standard Anthropic released in November 2024 to let agents connect to external services and data sources, has become unnecessary now that models are much stronger. The author recounts attending an all-day MCP-focused event and coming away tired of the protocol, calling it built for a time when LLMs were not that smart.

The post traces MCP's history: released by Anthropic in November 2024, it grew fast as users saw the productivity gains of giving models access to external tools, and Anthropic continued developing it before donating it to the Agentic AI Foundation, under the Linux Foundation, in December 2025. As adoption grew, users kept adding MCP servers, and each one's tools and schemas began overloading model context windows. Platforms including Composio, MintMCP, and Pipedream responded with generic search and execute patterns, giving agents one place to store credentials and a minimal tool set to reduce context bloat, which the author calls a good stopgap.

The author's central claim is that this tooling layer undersells how much models themselves have improved. Current models can execute code, reason over large codebases, and act more autonomously, including writing scripts that call APIs they have never seen before and composing multiple services into a workflow with little user intervention. The author also states that models have learned to run a CLI's --help command to discover its interface, so they often no longer need an MCP server to reach services that already expose documented APIs or CLIs, since most remote-service MCP servers just wrap existing APIs. As a case in point, the post cites Cloudflare's Code Mode, described in a Cloudflare Blog post by Kenton Varda and Sunil Pai published September 26, 2025, as a way to have LLMs compose calls into scripts executed in a sandbox rather than going through MCP tool calls one at a time.

The author's recommendation: delete most MCP servers, since agents with terminal access can replace them and are often more capable, while acknowledging that machine-readable CLI output such as JSON or XML can still be verbose and token-heavy, a problem the author says has workable fixes. Instead, the post argues for standardizing how agents use HTTP APIs directly, for instance by having agent clients attach identifying headers so servers can respond with Markdown or plain text instead of HTML or verbose JSON. Two examples are given: a growing number of LLM-friendly, text-heavy sites already honor an Accept: text/markdown header and return rendered Markdown instead of HTML; and a Vercel engineer proposed that agent harnesses send the client's preferred programming language via the Accept-Language header so documentation sites can serve language-specific examples, an idea Shopify's Tobi Lutke publicly endorsed and shipped in Shopify's own docs, replying "Great idea. Will support this on Shopify docs."

The post closes by framing MCP as a protocol of a bygone era: agents are now capable enough to write their own scripts and request exactly what they need, so the author argues it is time to end-of-life MCP and rely directly on HTTP APIs and CLIs wherever they already provide the needed interface.

Key facts

  • The post argues MCP, released by Anthropic in November 2024 and donated to the Agentic AI Foundation (under the Linux Foundation) in December 2025, is now obsolete because models have gotten far more capable.
  • Growing MCP adoption led to context bloat from many servers' tools and schemas; platforms like Composio, MintMCP, and Pipedream built generic search/execute patterns as a stopgap.
  • The author claims current models can execute code, call APIs directly, and use a CLI's --help command to discover its interface, reducing the need for MCP servers that just wrap existing APIs.
  • Cloudflare's Code Mode, described in a September 26, 2025 Cloudflare Blog post by Kenton Varda and Sunil Pai, is cited as a way to have LLMs compose calls into sandboxed scripts instead of individual MCP tool calls.
  • The post proposes agents identify themselves via HTTP headers so servers return Markdown or text instead of HTML/JSON; Shopify's Tobi Lutke publicly backed a related Accept-Language proposal from a Vercel engineer and shipped it in Shopify's docs.

Why it matters

The piece is a pointed bet against one of the more visible pieces of AI infrastructure built over the past two years. MCP became a de facto standard for connecting agents to external tools and services, and an entire secondary industry, monitoring, schema management, credential brokering, grew up around making it usable at scale. The author's claim is that this layer was solving a problem specific to weaker models, and that better models are quietly making the workaround unnecessary faster than the industry built out its tooling for it.

Who it affects

Anyone building or maintaining MCP servers or the platforms around them (the post names Composio, MintMCP, and Pipedream as examples of that ecosystem), teams that adopted MCP heavily and now face a decision about whether to keep investing in it, and API and documentation providers who could see agent traffic shift toward direct HTTP calls and CLI discovery instead of protocol-mediated access.

How to use it

The post is an opinion and a call to action rather than a shipped tool. Its concrete suggestions are: favor agents with direct terminal or API access over adding new MCP servers where a documented API or CLI already exists; if operating a documentation or content site, consider honoring an Accept: text/markdown header to serve agents rendered Markdown instead of HTML, and consider the Accept-Language proposal that Shopify has already adopted to serve language-specific code examples.

How solid is it

This is a single author's opinion piece with no data behind its core claims: no benchmark, cost figures, or token counts are given to support the assertions about context bloat or about how much better models have gotten at direct API and CLI use. The historical dates (MCP's November 2024 release, its December 2025 donation to the Agentic AI Foundation, Cloudflare's Code Mode post) are sourced to citations within the piece, and the Shopify anecdote is drawn from a public exchange, but the argument itself rests on the author's own experience and reasoning rather than measured evidence.

Risks and caveats

The post itself flags that CLI output in JSON or XML can be verbose and token-heavy, an issue it says has fixes without specifying them. It also does not name the 'big labs' it credits with being right about models improving, nor does it name or date the MCP-focused event that prompted the piece. The broader argument, that better models make a coordination protocol like MCP unnecessary, is contested territory: the same improvements it cites could equally be read as making tool protocols more useful rather than less, and the piece does not engage with that counterargument.

“Great idea. Will support this on Shopify docs.”

— Tobi Lutke (@tobi)