Agent Plugins 1.0: Why Claude Code Sat It Out

Agent Plugins 1.0 packages MCP into one portable format — but Anthropic, the protocol's own inventor, wasn't in the room.

7 min read

On August 6, 2026, six of the biggest names in AI tooling agreed on something so mundane it barely made headlines: a folder structure. Agent Plugins 1.0 is a coalition standard for packaging an AI agent’s Agent Skills and MCP servers into one portable directory — a plugin.json manifest, an optional skills/ folder, an optional mcp.json config — that any compliant client can install with a single command. It’s the kind of unglamorous plumbing decision that quietly reshapes an ecosystem, the way tar and gzip once did for Unix software distribution. I’ve spent enough years moving data through CERN’s grid computing pipelines and enough more racking servers in data centers to know a simple rule: whoever controls the packaging format controls the friction, and friction is where lock-in lives.

Here’s the thesis nobody else is writing about this launch: Anthropic invented the protocol this entire standard is built on. Model Context Protocol, MCP — the open spec that lets AI models call external tools and pull in live data — came from Anthropic, and Anthropic open-sourced it. Yet Anthropic was not part of the coalition that designed Agent Plugins 1.0. Amazon, Anysphere (Cursor’s maker), Microsoft, OpenAI, Vercel, and Google all signed on; Claude Code, Anthropic’s own coding agent, didn’t make the day-one support list and runs a completely different plugin format from the one the rest of the industry just agreed to share. Plugins built to the new spec still install into Claude Code today — this isn’t a boycott story — but the company that built the pipes wasn’t at the table when everyone else agreed on how to box up what flows through them.

What Agent Plugins 1.0 Actually Standardizes

Before Agent Plugins 1.0, if you built a useful capability — say, a code-review skill wired to a database MCP server — you packaged it differently for every client. Cursor wanted one layout, VS Code extensions wanted another, and each coding agent invented its own plugin conventions from scratch. That’s the exact fragmentation problem GitHub Copilot Code Review’s Agent Skills + MCP integration ran into when it tried to recentralize control over how skills and servers got wired together inside one vendor’s product.

Agent Plugins 1.0 fixes that by defining exactly three pieces in one directory: a plugin.json manifest describing the plugin’s identity and dependencies, an optional skills/ folder holding Agent Skills (packaged instruction sets an agent can load on demand), and an optional mcp.json file configuring the MCP servers the plugin needs to function. Write the folder once, and any client implementing the spec can install it. The spec itself is public at github.com/agentplugins/agent-plugins-spec, and the governance charter behind it is unusually strict for a six-company coalition: no single vendor is allowed to hold a majority of steering-committee seats, which is the kind of clause you write specifically because you’ve watched a “neutral” standard get captured before.

Who Showed Up on Launch Day

The first shipping implementation landed fast: OpenAI’s Codex CLI added support on August 7, 2026, in version 0.147.0, just a day after the spec published. That’s not a coincidence — when you help write the standard, you ship day-one, and Codex CLI now lets you search local, personal, workspace, and remote plugin catalogs and install straight from any of them. Five platforms had working support at launch: ChatGPT/Codex, Cursor, GitHub Copilot, Kiro, and VS Code. Claude Code is the visible gap in that list, and the chart below is not a hunch — it’s the actual day-one support matrix reported across coverage of the launch.

Agent Plugins 1.0: Day-One Platform Support Own format / not listed Supported at launch ChatGPT / Codex CLI Supported Cursor Supported GitHub Copilot Supported Kiro Supported VS Code Supported Claude Code Own plugin format — not on day-one list

Source: explainx.ai and releasebot.io coverage of Agent Plugins 1.0 launch, August 2026

Inside a Plugin: plugin.json, skills/, and mcp.json

Strip away the coalition politics and the mechanism is genuinely elegant. A plugin author drops three things into a single folder: the manifest declaring what the plugin is and what it depends on, a skills/ directory containing however many Agent Skills the plugin ships, and an mcp.json describing which MCP servers need to be running for those skills to actually do anything. That’s it — no per-client adapter layer, no separate build step for Cursor versus VS Code versus Copilot.

Compare this to how coding tools have historically handled distribution, where a platform’s own convenience layer becomes a quiet moat. Meta made a similar go-it-alone bet with its monorepo tooling strategy, discussed in Muse Code: Meta’s Monorepo Bet That Might Backfire — betting that owning the whole stack beats interoperating with it. Agent Plugins 1.0 is the opposite bet: six companies decided that a shared packaging format grows the whole pie faster than any one of them hoarding plugin compatibility. Codex CLI’s catalog system is the clearest proof this works in practice — it already searches local, personal, workspace, and remote plugin catalogs and installs directly, no manual file-copying required.

The Protocol’s Inventor, Missing From the Coalition

Here’s the part that should bug practitioners more than it seems to. MCP servers are the third leg of every Agent Plugin — the mcp.json file is meaningless without the protocol it configures — and MCP came from Anthropic. Anthropic built it, named it, and released it as open infrastructure specifically so the whole industry could build on it without asking permission. That’s exactly what happened here: five other major platforms used that openness to build a packaging standard on top of MCP, and Anthropic simply wasn’t part of drafting it.

To be precise, because this gets overstated everywhere else: Claude Code is not locked out. Plugins authored to the Agent Plugins 1.0 spec do install into Claude Code today. What Claude Code lacks is the write-once-run-everywhere convenience the standard promises everyone else — its plugin layout is its own, separate from plugin.json/skills/mcp.json, so a plugin author targeting the new spec still needs a Claude Code-specific path if they want full compatibility. It’s the practical equivalent of the company that invented USB not showing up to the meeting where everyone agreed on the connector shape — your devices still charge, technically, but you’re the one holding the adapter.

⚡ PHOTON’S TAKE

I’ve watched open protocols get adopted and then packaged by someone else more times than I can count — it’s how open infrastructure usually plays out once it actually works. MCP was always going to outgrow Anthropic’s direct control the moment it was open-sourced; that’s the whole point of open-sourcing something. What surprises me isn’t that six vendors built a shared packaging layer without Anthropic in the room. It’s that Anthropic let the meeting happen without them. Claude Code still installs these plugins fine today, but skipping the governance table on the format wrapping your own protocol is a strategic choice, not an oversight — and it’s the kind of choice that’s expensive to walk back later.

Where Agent Plugins 1.0 Goes From Here

Standards with a no-single-majority governance clause tend to have staying power, because no one vendor can quietly bend the spec toward its own roadmap later. Google’s own developer blog covering the launch is a tell here — when a company that could easily have built a competing proprietary format instead publishes explainer content pointing at a shared open one, that’s a signal the coalition expects this to stick, not fade in a quarter. The full spec is sitting in the open at github.com/agentplugins/agent-plugins-spec, which means Anthropic could, in principle, join the governance process tomorrow the same way any other vendor would.

My honest read: Anthropic will end up supporting Agent Plugins 1.0-compatible packaging in Claude Code within the next couple of release cycles, because the pressure from plugin authors who don’t want to maintain two formats will eventually force the issue. But “eventually” is doing real work in that sentence — every month Claude Code runs its own format while five competitors share one is a month where plugin authors default to building for the standard first and Claude Code second, if at all. The company that built the pipes doesn’t get to skip the plumbing meeting forever without the pipes starting to feel like someone else’s.

Photon Guy
Photon Guy

Photon Guy writes at the intersection of particle physics and heavy computing infrastructure. He spent years at CERN working on silicon particle detectors — the sensors that catch what the world's largest accelerators smash together — before moving into the data center industry, where he works on the machines that power the internet and AI. ScienceShot is where those two worlds meet: real physics, real engineering, strong opinions, and no press-release rewrites.

Articles: 33

Leave a Reply

Your email address will not be published. Required fields are marked *