Agent Tooling Churns. Your Files Don't Have To.

The way agents talk to your stuff gets rebuilt every cycle. A year ago the Model Context Protocol was the obvious bet. Now the discourse argues it is already dying, the command line is the future, and Skills are next. Each turn solves a real problem and introduces a new one. The file underneath does not move.

This is not a prediction that any one layer loses. The Model Context Protocol is the open standard that lets an agent reach into a file or call a tool, and by Anthropic's own December 2025 numbers it had "97M+ monthly SDK downloads across Python and TypeScript" and "more than 10,000 active public MCP servers, covering everything from developer tools to Fortune 500 deployments."¹ That is not a dead protocol. It is a layer being argued about — which is what every integration layer does on its way through the churn.

What everyone is arguing about

The argument is whether MCP is dying, and it runs in both directions at once. One camp says the protocol bloats the context window and the command line does the job better. The other says MCP is becoming boring infrastructure the way HTTP did. Both camps have a point. Neither settles the only question that touches your notes.

The "it's dying" case is not a troll. Eric Holmes, whose post set off much of the wave, wrote: "I'm going to make a bold claim: MCP is already dying."² His reasoning is mechanical, not contrarian. "CLIs compose. I can pipe through jq, chain with grep, redirect to files. This isn't just convenient; it's often the only practical approach."² A command line is something models are already fluent in. A protocol is one more thing to load.

There is a measurable failure underneath the complaint. A May 2025 paper by Tiantian Gan and Qiyao Sun found that large language models "struggle to effectively utilize a growing number of external tools" defined by MCP "due to prompt bloat and selection complexity."³ Their fix, retrieval, "more than triples tool selection accuracy (43.13% vs 13.62% baseline)."³ Read that direction carefully. Flooding the model with every tool definition drops selection to roughly fourteen percent; retrieving only the relevant ones recovers it to forty-three. The problem the churn is reacting to is bloat, not the file.

What the other side is right about

The steelman is that MCP is not losing. It is settling into the unglamorous role every standard eventually fills. Charles Chen, replying to the funeral, argued the opposite of decline: "MCP over streamable HTTP? This is an absolute game changer and will be a key linchpin in organizational and enterprise adoption."⁴ The protocol matures where the command line cannot follow.

Even the skeptics concede the same territory. Chloe Kim, the Quandri engineer behind a 400-point "MCP is dead?" thread posted in late May 2026,⁵ is blunt about the hype: "every SaaS landing page has 'MCP supported' in the feature list… Same pattern as 'AI-powered' and 'blockchain-based' marketing from years past."⁶ And yet she draws the line carefully. MCP is "still valid when" there is "no CLI" for a web-only service, or for "non-developer users" who do not live in a terminal.⁷ The honest reading is not a winner. It is a layer narrowing its scope while the next layer takes the rest.

So the discourse, across months and across both sides, has agreed on something quietly. The plumbing is in motion. What it plugs into is not the thing being debated.

The part nobody is fighting over

Notice what every voice in this argument takes for granted. Holmes pipes a CLI's output to a file. Kim wants a connection method for data she has. Chen wants enterprises to reach systems of record. Every position assumes a note already sitting somewhere — and argues only about the wire to it. The wire is contested. The thing on the other end is not.

That is the asymmetry worth keeping. The integration layer is where the energy goes, because it is where the differentiation and the funding are. MCP arrived, the command line answered, and Skills are the current challenger: a folder with one Markdown file that tells an agent how to do a task. Something will answer them in turn. Each layer is rebuilt because each one is, at the moment, annoying enough to replace. The file in an open format is never the thing anyone proposes replacing, because it is not anyone's product.

The velocity proves it from both ends. The protocol grew to ten thousand-plus servers in a year;¹ the CLI-and-Skills camp grew a single curated skills list to 13,621 stars in five months, created in January 2026.⁸ Two layers, opposite bets, both moving fast. The notes they reach for are written in the same plain text either one reads without conversion.

Why the file outlasts the plumbing

A plain file in an open format outlasts every integration layer for a structural reason, not a sentimental one: it has no vendor, no roadmap, and nothing to deprecate. MCP can be donated to a foundation, the CLI can fall out of fashion, Skills can be superseded, and the file reads exactly the same to whatever arrives next.

Migration is the cost the file does not charge. A Markdown note on your disk is read in place by today's tool and tomorrow's, with no conversion step in between.

The pattern is old enough to name. In 2019 Ink & Switch's local-first researchers (Martin Kleppmann, Adam Wiggins, Peter van Hardenberg, and Mark McGranaghan) gave it a motto: "You own your data, in spite of the cloud."⁹ The agent era only sharpens the point. When the tool that reads your data changes every cycle, ownership of the data is the one decision that does not expire. A note locked in a proprietary silo has to be migrated each time the plumbing turns over. A note in plain text on your own device is simply read in place, again, by the next thing.

This is why betting on a single layer is the wrong frame. The question is not MCP or Skills. It is whether the thing both of them reach for is yours, open, and readable without the app that made it.

What to do while the layer churns

Stop optimizing for the protocol of the month and optimize for the substrate underneath it. You cannot predict which integration layer wins next year, and you do not need to. You only need your notes to be in a shape that every contender already reads. Three moves follow, and none of them require adopting any specific tooling.

• Keep the notes you care about in plain text, in an open format. Markdown reads cleanly to a person and to any agent, with no export step between the file and the tool.
• Keep them on your own device, not locked inside one vendor's account, so a change in the plumbing never forces a migration of the content.
• Treat the integration layer as disposable — try MCP, try a CLI, try Skills — but never let the artifact you keep depend on which one you tried.

For the explainers behind the layers themselves, the companion pieces go deeper: what MCP actually is and how agents reach your files, and why an agent Skill is just a Markdown folder. This piece is about the part those two have in common: the file that outlives both.

Frequently Asked Questions

Is MCP dead? No. It is being argued about, which is different. By Anthropic's December 2025 figures MCP had "97M+ monthly SDK downloads" and "more than 10,000 active public MCP servers."¹ Its scope is narrowing — context bloat pushes some teams toward the command line and Skills³² — but a narrowing standard is not a dead one.

What is replacing MCP? Nothing wholesale. The pattern is that the integration layer keeps churning: MCP, then the command line, then Skills, then the next thing. Each handles part of the job. The honest answer is that no single layer replaces the others — they divide the territory and keep shifting.⁴⁷

Should I bet on MCP or Skills? Bet on neither exclusively. Both are integration layers that will be rebuilt or superseded. The durable bet is the substrate they both read — plain files in an open format on your own device — which survives whichever layer wins. Try the tooling freely; keep the artifact independent of it.

Why does MCP eat so much context? Loading many tool definitions at once floods the context window and degrades selection. Gan and Sun found this drops tool-selection accuracy to a 13.62% baseline, which retrieval recovers to 43.13%.³ The fix is feeding the model fewer, relevant tools — not abandoning the file the tools act on.

Is the agent tooling layer going to keep changing? Yes. The integration layer is where the competition and the funding concentrate, so it gets rebuilt on a cycle — MCP to CLI to Skills to next. The file format is the part that does not have to migrate, because it has no vendor and nothing to deprecate.

Did Anthropic give away MCP? Yes. In December 2025 Anthropic donated the Model Context Protocol to a vendor-neutral foundation under the Linux Foundation — the same announcement that reported the 97M-download and 10,000-server figures.¹ That made the standard no single company's property, but it did not stop the discourse about its scope.

The plumbing changes; the file stays. Whatever the agent of the month turns out to be, it reads the notes you already have. If those notes are plain text on your own device, mnmnote.com keeps them that way, in your browser, in a format anything can open.