Carry Your Plugins Across Agents — Codex 0.147 Is the First to Ship the Shared 'Agent Plugins' Standard, and the Empty Seat Where Claude Should Be

You built a skill or set up an MCP configuration, and you'd like to reuse it in a different coding agent — but until now, every tool has had its own location and its own format for these things. You had to rebuild the same extension, rearranged for each client. A shared standard aiming to kill that hassle started moving in August.

Released on August 7, 2026, Codex CLI 0.147.0 became the first major agent to support that standard, called "Agent Plugins." The day before, on August 6, the standard itself was announced simultaneously by several companies as a vendor-neutral standard. This time we'll walk through both of these developments from a hands-on developer's point of view.

What Codex 0.147 Ships

The headline of 0.147.0 is that it moves extensions from "holding them in a config file" to "finding and installing them as distributed artifacts." Since there are several changes, here are the main additions.

• Agent Plugins install and cross-catalog search: Search for and install plugins across local, personal, workspace, and remote catalogs.
• --approve-for-me flag: Clears approvals via a single automatic-review step instead of an interactive prompt. The aim is to drop the confirmation dialog in unattended runs and CI while keeping the sandbox's containment intact.
• Persistent sections for conversations: Group threads into manually arranged sections so you can work through long histories a piece at a time.
• Opt-in support for MCP 2026-07-28: Supports paged discovery, multi-stage requests, and non-blocking server startup, letting you read the tool list without waiting for the server to initialize.
• Amazon Bedrock improvements: Adds cached web search and remote conversation compaction.

The thread organization and import features are a continuation of what began in 0.145/0.146 at the end of July. What's new this time is that it hands the very way you install extensions over to an external, shared standard.

What Is 'Agent Plugins 1.0'?

Agent Plugins 1.0 is a vendor-neutral specification that unifies the "container" for agent extensions. The core is almost disarmingly simple: a plugin is defined as a single directory. Its contents are these three elements.

What matters is that this standard doesn't invent a new format for the parts themselves. Skills still follow the separate Agent Skills spec and MCP still follows the MCP spec; the standard stays a thin layer that simply "wraps them without rebuilding them." A single install lets you distribute skills and MCP servers together.

The goal is to end fragmentation. Until now, VS Code listed four formats side by side — Agent Plugins, Copilot, Claude, and the old OpenPlugin — while Cursor and Codex kept their own separate locations in .cursor-plugin / .codex-plugin. The idea is that by settling on one shared lowest common denominator, the same parts can move between clients without any rewriting.

Who Runs It, and Who Isn't There

Governance is designed not to lean on any single company: the technical steering committee is made up of five people. The charter stipulates that "no single vendor can hold a majority of the core-maintainer seats," and the assets are held in trust by a neutral entity.

Google, too, announced it would join as a core maintainer on the day of the launch. The clients supported at launch are listed as ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code.

What stands out is that Anthropic's name isn't here. Claude Code uses its own layout at the hidden path .claude-plugin/plugin.json, which doesn't mesh as-is with Agent Plugins' convention of placing plugin.json at the root. One commentary reads this gap as "more unfinished than a rejection," but the fact that Claude — a heavy hitter among coding agents — isn't in a governance seat is worth noting as a difference in temperature over who leads the standard.

What It Means on the Ground, and the Caveats

From a team-operations standpoint, this pushes "extension standardization" one step further. If you can distribute the same skill or MCP configuration in the same container across Codex, Cursor, and Copilot alike, then switching tools is less likely to waste the assets you already have. For organizations that want to avoid lock-in, it's a tailwind.

On the unattended-operations side, --approve-for-me reduces friction in CI and headless runs. But automating approvals is also an operation that removes the checkpoint of interaction. The sandbox containment reportedly remains, but since you're skipping the human step of "stopping to confirm," the operator needs to draw a line on how far things may be auto-approved. It's reasonable to keep the concern — that checks thin out behind the convenience — in the back of your mind.

The standard itself comes with caveats, too. The spec page proclaims version 1.0.0, yet its status is "Working Draft," and no 1.0.0 git tag or release has been confirmed yet. In other words, the direction is set but the details are still being nailed down. Rather than rushing a full migration right now, the right distance seems to be this: watch how the roster of supported clients grows and whether Claude Code moves toward this shared format, while getting your existing skills into a shape you can port.

References: ChatGPT & Codex changelog (0.147.0) / openai/codex Release 0.147.0 / Agent Plugins 1.0: What the Standard Actually Fixes (digitalapplied) / Agent Plugins: the new open standard for AI agent extensions (eesel AI) / Codex Updates - August 2026 (Releasebot)