May 29, 2026
Plug-in panic in chatbot land
MCP Is Dead
Devs declare AI’s favorite plug-in overhyped as commenters drag the “dead” claim
TLDR: A new post says the tool system used by many AI assistants is slow, unreliable, and wastes valuable chat space, though one major memory problem has already been reduced in newer versions. In the comments, people split hard between “MCP is overhyped” and “this takedown is outdated,” turning the debate into a full-blown nerd brawl.
The article came in swinging: a backend engineer argued that MCP — the system that lets chatbots connect to apps like Slack, Notion, and Linear — is basically more trouble than it’s worth. Her case was blunt: it hogs space, breaks too often, runs slower than going straight to the app, and duplicates tools developers already have. The spiciest stat? In their setup, just describing the available tools ate up 10.5% of Claude’s working memory and 16.5% of GPT-4o’s. That’s the kind of number guaranteed to start a comment-section food fight.
And wow, the crowd did not quietly agree. One camp yelled, essentially, “This article is already outdated!” because Claude has since rolled out on-demand loading that reportedly cuts that memory bloat by 85%+. Critics pounced on the piece with eye-rolls and sarcasm, with one calling it part of the wave of “AI slop articles about AI” and another dunking on the idea that a command line magically avoids extra layers. But the anti-MCP crowd wasn’t exactly shy either: some said they’d been burned early and still prefer simple scripts, calling them leaner, cheaper, and easier to control. The most reasonable middle-ground take? MCP may still make sense at the company-wide level, especially when non-technical staff need safe access to internal tools. So no, the internet did not hold a funeral — it held a messy family argument, complete with nitpicks, nostalgia, and a lot of “actually...” energy.
Key Points
- •The article argues that MCP has three main drawbacks in practical use: high context-window consumption, lower operational reliability, and overlap with existing CLI/API approaches.
- •Quandri measured 77 MCP tools across Linear, Notion, Slack, and Postgres at roughly 21,077 tokens total, which the article says equals 10.5% of Claude's 200K context window and 16.5% of GPT-4o's 128K window.
- •Linear alone accounted for about 12,807 estimated tokens in the article's measurements, making it the largest contributor among the measured MCP servers.
- •The article cites a referenced benchmark claiming Jira MCP was 3x slower per call than direct REST API access and 9.4x slower on first call including initialization.
- •An update added to the article says Claude Code's Tool Search with Deferred Loading reduces MCP tool-schema context usage by more than 85%, largely addressing the article's original context-bloat concern.