MCP is dead, apparently
The Perplexity CTO announced at their developer conference this week that they're moving away from MCP internally. Garry Tan tweeted "MCP sucks honestly." Pieter Levels called it useless. The "MCP is dead, long live the CLI" post hit the top of Hacker News. OpenClaw, the hottest open-source agent project in the world, deliberately chose not to support it.
The argument: MCP tool definitions eat your context window. Auth is clunky. The whole thing is an unnecessary abstraction over APIs that already exist. LLMs are smart enough to call APIs directly, or use CLIs, or write their own integration code. Why add a protocol layer?
This sounds right if you're a developer wiring up tools on your laptop. It falls apart the moment you try to run agents across a team.
The context window complaint is real. Loading 50 tool schemas into context before the agent reads your prompt is wasteful. But this is a solved problem. OpenAI just shipped tool search in their Responses API, where tools only load when the model needs them. Anthropic and Cloudflare independently converged on "Code Mode," where agents write code against MCP tools instead of calling them through context. Anthropic tested it and cut token usage by 98.7%. The context window problem is last year's problem.
The CLI argument is more interesting. "Just give the agent a CLI and it'll figure it out." This is true. Agents are great at CLIs. But a CLI runs as you. Your credentials, your permissions, no distinction between "I did this" and "my agent did this." For one developer on their own machine, fine. For a company with 50 engineers and agents touching production systems, terrifying.
MCP creates a permission boundary that CLIs don't have. The spec mandates OAuth 2.1 for remote servers. You get user-level consent, token expiry, scoped access, audit trails. The CLI alternative is API keys in environment variables. Yikes.
Even the projects being held up as proof that MCP lost tell the opposite story. Perplexity still runs an MCP server for developers. Their new Agent API is basically a managed MCP gateway with better branding. OpenClaw's skill marketplace, ClawHub, runs on MCP servers underneath. Cloudflare's Code Mode, framed as the "MCP killer," literally uses MCP's discovery layer to find tools before converting them to a typed API. These projects didn't abandon the protocol. They wrapped it in better DX.
17,000 MCP servers. 97 million monthly SDK downloads. Stewarded by the Linux Foundation. Adopted by OpenAI, Google, Microsoft, and AWS. The 2026 roadmap is focused on horizontal scaling and enterprise auth. This is not a dying ecosystem.
MCP isn't dying. It's just becoming infrastructure. And infrastructure is boring. Nobody writes viral tweets about TCP/IP.
Replies
Basedash
For sure. Every protocol goes through this. REST was "unnecessary" when we had SOAP. OAuth was "too complicated" when we had API keys. The complaints are always about developer ergonomics on day one, never about what happens at scale on day 1,000.
The CLI argument breaks down exactly where you described — it doesn't draw a line between "I did this" and "my agent did this." That distinction doesn't matter for one dev on their own machine. It matters enormously the moment you have audit requirements, shared production access, or more than one person who can trigger an agent.
The reason the "MCP is dead" takes keep spreading is the same reason good infrastructure stays invisible: the projects shipping on top of it don't lead with "powered by MCP." Perplexity doesn't market their Agent API as a managed MCP gateway. OpenClaw doesn't advertise ClawHub as running on MCP servers underneath. They market the capability. The protocol disappears into the stack — and that's actually the sign it's working, not dying.
I don't know if I buy this.
The basic premise of MCP is - if you design a tool for agents, it will be better than non-agent tool.
Another example: Computer use.
Yes, agents will eventually gain human-tier computer use. Is anyone under the illusion that they will be more productive doing computer use than they are working programmatically via code/CLI?
Similarly, I'm certain that programmatic agent-designed interfaces like MCP will outperform legacy interfaces.
The challenge is that many MCPs are overcrowding context and there is no good system to manage this at present. So there is a tension right now between the desire to provide AI with all the possible tools and manage context/performance. I assume this tension will relax over time as infrastructure improves, MCPs become more efficient, and context size increases.
Just some thoughts off the top of my head. I also think it's possible even better interfaces will emerge that surpass MCP.