The MCP backlash finally found its argument
Maharshi Patel's post arguing MCP was always a bad idea landed on Hacker News with 316 points and 303 comments, which is the ratio of a fight rather than a consensus. The argument deserves engaging with because it is not the usual complaining about tool schemas.
The core claim is that MCP was designed for a model capability level that no longer exists. Anthropic released it in November 2024, when models were, in Patel's words, relatively primitive, and needed a structured protocol to be told what a service could do. Current models execute code, reason about large codebases, and call APIs they have never seen before. They discover CLI interfaces by running the help flag. The scaffolding was built for a model that needed scaffolding.
The practical complaint is context bloat, and this one is not arguable. Every MCP server you connect contributes a pile of tools, each with a schema, all of it sitting in the context window before the model has done anything. Connect five servers and a meaningful fraction of your window is spent describing capabilities the model will not use on this task. Cloudflare's Code Mode is the counterexample he leans on: let the model compose calls into a script and run it in a sandbox, which is both more expressive and cheaper than round-tripping through tool definitions.
His proposal is where it gets interesting, because it is small and shippable rather than a new standard. Make HTTP APIs agent-aware with headers. Accept markdown for documentation-heavy sites so the agent gets prose instead of a div soup. Accept-Language specifying a programming language preference, which Shopify has already adopted. Use the protocols that built the internet instead of inventing one.
Where I think the post oversteps: MCP is not only a capability description, it is also an authorization and connection boundary, and give-the-model-a-shell-and-credentials answers the capability question while quietly making the security question much worse. The strongest version of Patel's argument is not that MCP should die, it is that MCP should be lazy, with tools discovered on demand rather than dumped into context up front. That is a fixable design flaw being argued as an original sin.
← Back to all articles
The core claim is that MCP was designed for a model capability level that no longer exists. Anthropic released it in November 2024, when models were, in Patel's words, relatively primitive, and needed a structured protocol to be told what a service could do. Current models execute code, reason about large codebases, and call APIs they have never seen before. They discover CLI interfaces by running the help flag. The scaffolding was built for a model that needed scaffolding.
The practical complaint is context bloat, and this one is not arguable. Every MCP server you connect contributes a pile of tools, each with a schema, all of it sitting in the context window before the model has done anything. Connect five servers and a meaningful fraction of your window is spent describing capabilities the model will not use on this task. Cloudflare's Code Mode is the counterexample he leans on: let the model compose calls into a script and run it in a sandbox, which is both more expressive and cheaper than round-tripping through tool definitions.
His proposal is where it gets interesting, because it is small and shippable rather than a new standard. Make HTTP APIs agent-aware with headers. Accept markdown for documentation-heavy sites so the agent gets prose instead of a div soup. Accept-Language specifying a programming language preference, which Shopify has already adopted. Use the protocols that built the internet instead of inventing one.
Where I think the post oversteps: MCP is not only a capability description, it is also an authorization and connection boundary, and give-the-model-a-shell-and-credentials answers the capability question while quietly making the security question much worse. The strongest version of Patel's argument is not that MCP should die, it is that MCP should be lazy, with tools discovered on demand rather than dumped into context up front. That is a fixable design flaw being argued as an original sin.
Comments