Skills, Plugins, and Marketplaces Across Coding Agents (May 2026)

A practical survey of how coding agents other than Claude Code package reusable behavior, and how portable those formats actually are.

Reference: Claude Code's "Agent Skills" are folders with a SKILL.md + YAML frontmatter (name, description), auto-loaded from ~/.claude/skills/ or .claude/skills/. Plugins use .claude-plugin/plugin.json and are installed from a marketplace (.claude-plugin/marketplace.json) via /plugin marketplace add and /plugin install. Anthropic published this as an open Agent Skills spec.

The big picture in 2026: the spec has largely won. Most agents now either load Anthropic-style SKILL.md natively or via a thin shim, and AGENTS.md + MCP are the two cross-agent standards everyone agrees on.

1. Gemini CLI (Google)

Equivalent: "Extensions" (the umbrella) and "Agent Skills" (introduced March 2026).
On-disk format: Extensions live in ~/.gemini/extensions/<name>/ with a gemini-extension.json manifest. They can bundle MCP servers, GEMINI.md context, custom slash commands as .toml, hooks, sub-agents, themes, and skills. Skills go in <extension>/skills/<name>/SKILL.md, or directly in ~/.gemini/skills/ / .gemini/skills/ for ad-hoc use (Gemini CLI extensions, reference).
Install / distribution: gemini extensions install <github-url-or-path> clones into the extensions dir; gemini extensions update re-pulls. There is an official browseable catalog at geminicli.com/extensions and a gemini-cli-extensions GitHub org (github.com/gemini-cli-extensions).
Skill-like unit can contain: prompt + optional supporting files; in extensions, also tools (via MCP), hooks, sub-agents.
Spec compatibility: The skills format is deliberately Anthropic-compatible — same SKILL.md + YAML frontmatter (Google Cloud Community: "Agent Skills"). A community shim, gemini-cli-skillz, already exposed Anthropic-style skills via MCP before native support landed.

2. OpenCode (sst/opencode)

Equivalent: "Agent Skills" (native) + "Plugins" (TypeScript hooks).
On-disk format: SKILL.md files with YAML frontmatter (name, description), under repo .opencode/skills/ or ~/.config/opencode/skills/ (Agent Skills | OpenCode). Plugins are TS modules registered in config and exposing lifecycle hooks (Plugins | OpenCode).
Install / distribution: Mostly git clone / npm today. There is no official first-party marketplace UI, but several third-party catalogs exist: awesome-opencode, n-skills (curated, multi-agent), and BuySkills (paid + free).
Skill-like unit: prompt + supporting files; plugins add tools/hooks. MCP is a first-class tool surface.
Spec compatibility: Native Anthropic Agent Skills — same SKILL.md shape, intentionally cross-agent. Drop-in compatible with Claude Code skill folders in practice.

3. Codex CLI (OpenAI)

Equivalent: AGENTS.md for always-on instructions, Agent Skills for on-demand workflows, and Plugins as a packaging layer.
On-disk format: AGENTS.md is plain markdown, layered from repo root downward, concatenated nearest-last so deeper files override (Codex AGENTS.md guide). Skills are SKILL.md folders, mirroring Anthropic's spec (Agent Skills – Codex). Custom prompts are deprecated in favor of skills.
Install / distribution: Skills can ship inside plugins, which can also bundle "app mappings", MCP server config, and presentation assets in a single package. Distribution is git-based; no centralized marketplace UI as of May 2026.
Skill-like unit: instructions + optional scripts/resources; plugins add MCP servers and tool config.
Spec compatibility: Native Anthropic Agent Skills. Codex was the first non-Anthropic agent to commit to AGENTS.md, and the second (after Gemini) to ship the skills spec natively.

4. Cursor

Equivalent: Rules (.cursorrules legacy, .cursor/rules/*.mdc modern), MCP for tools, plus Agent Skills (added 2026).
On-disk format: .cursor/rules/*.mdc are markdown-with-frontmatter rule files, scoped per-glob and either always-on or auto-attached (Rules | Cursor Docs). Skills follow SKILL.md (Mastering Cursor 2026).
Install / distribution: Rules are manual copy / git; the community hub is cursor.directory (no in-IDE installer until recently). MCP servers install via Cursor's settings UI.
Skill-like unit: Rules are prompt-only. Skills can include workflows, hooks, and pull in domain knowledge on demand. Tools come via MCP, not via the rules format.
Spec compatibility: Cursor reads AGENTS.md and now Anthropic-style SKILL.md folders (awesome-cursor-skills). Rules → skills migration is encouraged for richer workflows.

5. Aider

Equivalent: Conventions files (any markdown, loaded with --read) and built-in slash commands. No plugin/skill system.
On-disk format: Plain markdown convention docs (often CONVENTIONS.md); .aider.conf.yml for config (Aider docs).
Install / distribution: Manual. Aider is intentionally minimal — no marketplace, no plugins.
Skill-like unit: Effectively a system-prompt fragment. No tool/hook/subagent extensibility beyond what aider builds in.
Spec compatibility: Reads AGENTS.md if you point it at one with --read AGENTS.md. Does not auto-discover SKILL.md. Closest cross-agent surface is "write your conventions in AGENTS.md and --read it."

6. Cline / Roo Code

Cline: Custom instructions file, .clinerules, and a first-party MCP Marketplace inside the VS Code extension — one-click install of MCP servers (Roo vs Cline 2026). No native Anthropic-style skills.
Roo Code: Forked Cline. Adds Custom Modes — named personas (Architect, Debugger, ESP32 Expert, etc.) defined in .roomodes JSON, with a Custom Mode Marketplace / Mode Gallery for sharing. Supports MCP but no MCP marketplace (Roo Code docs).
- Note: Roo Code announced shutdown of all products on May 15, 2026 (Morph comparison); the team is migrating to roomote.dev. Treat Roo formats as legacy.
On-disk format: .clinerules (markdown), .roomodes (JSON of mode definitions).
Skill-like unit: A "mode" is system prompt + tool allowlist + model config — closest non-Anthropic analog to a skill but agent-bound, not a portable folder.
Spec compatibility: Neither natively loads SKILL.md. Both speak MCP.

7. Continue.dev

Equivalent: config.yaml (formerly config.json), custom slash commands, rules, and the Continue Hub (rebranded "Mission Control").
On-disk format: ~/.continue/config.yaml declares models, context providers, slash commands (named prompt templates), rules, and MCP servers (Continue config reference).
Install / distribution: Hub / Mission Control is a real web marketplace — browse configs, one-click add, auto-sync across IDEs, share with team (Hub vs Local). This is the most polished marketplace UX outside Claude Code.
Skill-like unit: Slash commands are prompt + optional model routing; rules are always-on context; MCP supplies tools. No equivalent of a self-contained skill folder with hooks + scripts.
Spec compatibility: Does not natively load SKILL.md. You'd reimplement a skill as a custom slash command + rule + MCP server.

AGENTS.md as the de-facto cross-agent standard

AGENTS.md is the closest thing to a universal coding-agent format today. agents.md reports 60,000+ projects using it, and it's read natively by Codex, Cursor, Copilot, Gemini CLI, OpenCode, and Aider (via --read). It's now stewarded by the Agentic AI Foundation under the Linux Foundation, alongside MCP and the Goose framework (AGNTCon + MCPCon 2026). Claude Code reads CLAUDE.md but most teams symlink or duplicate to AGENTS.md (AGENTS.md vs CLAUDE.md).

It is prose-only — no tools, no hooks, no subagents. It's the lowest common denominator, and that is its strength.

MCP as the cross-agent tool surface

Every agent in this list speaks MCP: Claude Code, Codex, Gemini CLI, OpenCode, Cursor, Cline, Roo, Continue. If you want to ship reusable tools (not instructions) to N agents with one artifact, MCP is the only realistic answer in 2026.

What's actually portable today?

A realistic verdict:

Artifact	Portability
MCP server	Universal. Works in all 8 agents surveyed. Best surface for tools.
AGENTS.md	Near-universal for prose instructions. Codex, Cursor, Gemini, OpenCode, Aider, Copilot.
Anthropic `SKILL.md` skill	Native in Claude Code, Codex, Gemini CLI, OpenCode, Cursor. Not native in Aider, Cline, Roo, Continue — but most can be shimmed via MCP. This is the new emerging standard for richer-than-prose units.
Slash commands	Per-agent. Claude `.md` ≠ Gemini `.toml` ≠ Continue YAML ≠ Cursor. Don't try.
Hooks	Per-agent. Claude, OpenCode, Gemini, Codex each have their own hook shapes.
Subagents / modes / personas	Per-agent. Claude subagents, Roo modes, Cline modes, Codex subagents — different file shapes, no convergence yet.
Plugin / marketplace manifest	Per-agent. Claude `plugin.json`, Gemini `gemini-extension.json`, Codex plugin, OpenCode plugin — not interchangeable.

Practical recipe for a cross-agent capability in 2026:

Ship tools as an MCP server (npm or container).
Ship instructions as AGENTS.md at repo root, plus a Claude-compatible SKILL.md folder for richer on-demand workflows (works natively in Claude/Codex/Gemini/OpenCode/Cursor).
Maintain a thin per-agent plugin manifest (.claude-plugin/plugin.json, gemini-extension.json, etc.) that points at the same skills + MCP server. The bodies are reused; only the wrapper differs.

So: one artifact for everything is not realistic, but two artifacts (MCP + SKILL.md) plus thin per-agent manifests covers the vast majority of the modern coding-agent landscape.

Caveats: the skill ecosystem moved fast in Q1–Q2 2026 (Codex deprecated custom prompts in April; Roo announced shutdown for May 15; Gemini's skills support is March-2026 new). Verify the latest docs before committing to any format.

12 KiB Raw Permalink Blame History Unescape Escape