Maharshi Patel argues teams should delete most MCP servers and let agents run CLIs instead: Reveneau

Why it mattersIf the argument holds, the MCP servers, gateways and monitoring layers many teams have built over the last year become work to remove, and the default plan for the next tool becomes a documented CLI or HTTP API the agent can reach on its own.

An essay by developer Maharshi Patel titled "Why MCP Was Always a Bad Idea" reached the front page of Hacker News on 20 September with 319 points. Patel's argument, published on his personal blog on 14 September, is that the Model Context Protocol was designed for a moment in late 2024 when models could not reliably call arbitrary services on their own, and that models are now good enough that most MCP servers should be deleted.

The moment MCP was built for

Patel writes that Anthropic released MCP in November 2024, when Claude Code did not exist and general agentic workflows were "far less reliable". Adoption grew alongside the growth of LLMs across the economy. MCP was donated to the Agentic AI Foundation, under the Linux Foundation, in 2025. As users stacked more MCP servers, Patel says, the community started running into context bloat, where every server added its own tools and schemas to a model's context window.

What agents can do now

Patel says the models have moved on. They can run scripts, reason about large codebases, and compose calls to services they have never seen before. He cites Cloudflare's Code Mode, published on the Cloudflare blog on 26 September 2025, in which agents write TypeScript that composes several MCP calls and runs the whole script inside a sandbox in one pass. He also argues that agents with terminal access can now discover a CLI's shape by running it with the --help flag, so an MCP wrapper around a service that already has a CLI is doing work the agent could do on its own.

The proposed alternative

Patel's recommendation is short: delete most MCP servers, and standardise how agents talk to HTTP APIs directly. Two content-negotiation headers already work today, both of which he documents. Servers that support the Accept: text/markdown header can send a rendered Markdown version of a page, which several documentation sites already honour. Vercel engineer Malte Ubl proposed on X that harnesses also send an Accept-Language header naming the programming language the user wants, so a docs site can prioritise the Python example over the generic one. Shopify's Tobi Lutke replied on the same post that Shopify docs would support it.

Patel is explicit about what stays. Documented HTTP APIs, standard content negotiation, and mature authentication mechanisms already exist. The task, he writes, is to standardise how agents use those, not to build a parallel protocol around them.

Nothing in the piece measures how many current MCP servers this argument would delete, so a team reading it has to make the call server by server. The clear cases in his framing are MCP servers that only wrap a public API a coding agent could reach with curl and a documented endpoint. The remaining ones, in his framing, are servers that expose something an agent cannot reach any other way. For anyone building the next MCP wrapper for a public service, this is now the argument to answer, and it sits at 319 points on Hacker News.

Source

Maharshi Patel, Why MCP Was Always a Bad Idea, 14 September 2026.
Kenton Varda and Sunil Pai, Code Mode: the better way to use MCP, Cloudflare Blog, 26 September 2025.
Hacker News discussion, 20 September 2026 (319 points).

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.