GitHub Just Made Agent Plugins Portable — What Agent Builders Should Do Next

GitHub published Agent Plugins 1.0 on August 12, turning a familiar frustration in AI coding into a standardization problem with a practical answer: package the skill and the MCP server together once, then let compatible agent clients discover the parts they support. GitHub says the open standard is generally available in VS Code, Copilot CLI, the Copilot SDK, and the Copilot app, with support from AWS, Anysphere, Microsoft, OpenAI, and Vercel. Read the original release.

What You'll Learn

Why portable packaging matters more than another model leaderboard; what belongs in skills/, mcp.json, and the com.github.copilot/ namespace; how enterprise managed settings and MCP allowlists change the review boundary; and how to turn one repeatable agent workflow into something you can carry between tools without treating portability as automatic permission.

Step 1: Think in workflows, not prompts

A useful plugin is a repeatable outcome: release triage, deployment checks, test diagnosis, or documentation updates. Write down the trigger, the required context, the tools, and the review step before you package anything. AgenticNode is a useful place to study this orchestration mindset.

Step 2: Separate the portable core

Agent Plugins 1.0 standardizes skills under skills/ and MCP configuration in mcp.json. Client-specific agents, commands, rules, and hooks can live under com.github.copilot/, which other clients can ignore. That lets one package stay portable without pretending every client behaves identically.

Step 3: Keep the permission boundary visible

A portable MCP server is still a server that can read data or take actions. GitHub’s release points to managed settings such as enabledPlugins, extraKnownMarketplaces, and strictKnownMarketplaces, plus MCP allowlists that can approve or block servers by URL, command, or name. Use those controls, and test what happens when a plugin is denied.

Step 4: Migrate on a branch

Add the $schema field to plugin.json, move shared assets into the standard directories, and keep your old client-specific package until each client passes the same acceptance checklist. Review the diff, run the workflow with harmless fixtures, and verify that a human approval still gates every write or external message. The Vibe Coding Ebook now tracks the migration shape in Chapter 5.

Conclusion

Agent Plugins 1.0 is a small packaging change with a large workflow implication: skills and tools are becoming reusable infrastructure that can move across editors, CLIs, and agent surfaces. The winning habit is to standardize the workflow while keeping permissions, provenance, and review explicit. For the practical weekly version of this shift, subscribe to the Vibe Coding Academy newsletter.