Codex's Real Ambition: What MCP and Skills Still Can't SolveCodex 的野心，MCP 和 Skill 的下一步

Lately I've been using Agent apps like Codex App and Cursor intensively, and one thing has been striking me as more and more interesting.

Last year everyone was arguing about whose model was stronger; this year the argument seems to have shifted to whose right-hand panel is more usable.

Codex, the Claude desktop app, Cursor 3.0, TRAE SOLO — these top-tier Agents, without any coordination whatsoever, have almost simultaneously converged on the same interface layout: a list of projects and sessions on the left, a conversation with the Agent in the middle, and on the right a workspace holding file browsing, web preview, and diff review.

This definitely isn't a case of copying each other — it looks more like the current optimal solution for Agent interaction.

Why three columns

A traditional chatbot only needs two columns: conversation history on the left, the chat window on the right — you ask, it answers, and you leave when you're done.

In the Agent era, the Agent can write code, edit files, and call tools on its own. After it finishes, you need to check whether it did things right — the right-hand workspace exists precisely for that.

But this is only the first stage.

As users spend more and more of their time directing Agents, the time spent opening professional tools like VSCode naturally keeps shrinking. That question was bound to surface eventually: after the Agent finishes writing your code or your slide deck, if you want to tweak a few words, do you really have to switch out to a separate piece of software?

Nobody wants that. The user's natural expectation is: can't I just edit it directly inside the Agent? This is also one of the most requested features for Codex App right now (another is a mobile version, which is coming soon).

So each company quietly started upgrading its right-hand workspace, turning it from something that could only show file edit history into a multi-purpose panel. In Codex's April 16 major update, the changes to the right-hand workspace were the biggest of any feature.

The interaction details differ slightly between them. Codex and Cursor use tabs to switch; Claude uses a floating panel. Having used all of them myself, I find Codex the most comfortable — Claude's floating-panel design has more style than substance, and it's bound to change eventually.

Codex's real ambition

But if you read this change merely as "interface design evolving," you're underestimating Codex.

When Codex's big April update launched, its slogan was "Codex for (almost) everything" — capable of handling almost any task. You could read that as just an advertising slogan, but it's more like a statement of product direction.

To make good on that statement, Codex can't just be an Agent that's good at writing code — it has to handle all kinds of file formats, support professional workflows across every domain, and let users complete the entire loop within it, including the final manual touch-ups.

Right now Codex still can't do that last step: once something is generated, it can't be edited — not code, not Markdown, not PPTX. This might be a deliberate restraint on the product side, or it might be that the technology isn't there yet, or it might be waiting for a unified solution to emerge.

My guess is the third option.

MCP and Skills each solve only half the problem

To understand what Codex is waiting for, we first need to figure out which piece is still missing from the Agent capability puzzle.

MCP solves the "connection" problem: through a unified spec, the Agent can plug into all kinds of tools — databases, calendars, code repositories — all of it interoperable.

Agent Skills solve the "how to do it" problem: the Agent learns domain knowledge and best practices it wasn't trained on, like how to write in a particular style, or how to handle a certain kind of complex task.

Both of these are done fairly well. But there's one gap that's never been filled: the user's second-pass editing.

You have AI write you an article, and in the end you still have to open an editor yourself and tweak a few spots — after all, a lot of the time that last 5% of precision can only be nailed by doing it yourself. Even if AI gets smarter in the future, it will never understand you a hundred percent, so manual edits will always be needed.

So lately Markdown editors have become hot again, with all sorts of vibe-coded Markdown products popping up everywhere.

But Codex won't build its own Markdown editor, because everyone's preferences differ — whatever you build, someone will always be unhappy with it; and besides, it's simply impossible to integrate a specialized editor for every single vertical domain.

The most sensible path is a plugin mechanism.

Next step: an App Store for Agents

Turn the Agent into a platform and let the community contribute plugins — just like VSCode and Chrome do.

Codex only needs to focus on the Agent-orchestration layer, and hand file preview, second-pass editing, and vertical-domain expertise off to plugins to extend. Users install what they need — designers install design plugins, writers install writing plugins.

The plugin mechanism could also incidentally solve a problem that's had no answer for a long time: Skills can't be monetized.

My own baoyu-skills is nearing 20,000 stars, but I've made exactly $0 from it. A Skill is almost completely transparent — transparent to the Agent, and transparent to people too — the cost of copying it is essentially zero, so no matter how well you write it, the moat is shallow.

Plugins are different. The App Store and the Chrome extension marketplace have already proven out a working system for charging money and protecting copyright, and it's entirely feasible to transplant that to an Agent plugin marketplace. Good plugins could charge money, which gives developers the motivation to keep refining them, and only then does the ecosystem actually start turning.

Codex already has a very primitive plugin marketplace now. There's still a long way to go from here to a mature, paid plugin ecosystem, but the direction is right.

Codex isn't the only one trying to do this. I can see a similar shadow of it in Cursor. The only ones showing no sign of moving in this direction so far are Claude Code and Cowork — maybe they don't think it's worth doing, or maybe they just haven't gotten there yet.

A window left open for small and mid-sized teams

If Codex really does pull off a plugin ecosystem, what would that mean for small and mid-sized teams?

Besides building your own vertical Agent, there's another path: build plugins on top of a platform like Codex. You wouldn't need to build your own Agent-orchestration layer, wouldn't need to solve token access, and user distribution would ride on the platform too. All you'd need to focus on is that "last mile" — helping users take what the Agent generates and process it, polish it, and make it comfortable to use.

This window won't stay open very long. Those who get in early can capture the cold-start dividend; those who arrive late will be left fighting over scraps.

The timing isn't far off — maybe within just the next few months.

Codex's ambition is right there in front of us. For the slogan "almost any task" to actually come true, a plugin mechanism is a step that can't be skipped. If OpenAI keeps hesitating on this, that would be a genuine mistake.

Which company do you think will be the first to pull off this plugin ecosystem? Or do you think there's a product form even better suited to Agents? Feel free to share your thoughts in the comments!