Skills, MCP servers or plugins: which one, and where each lives

A coding agent has seven extension points: AGENTS.md (standing rules), skills (procedures loaded on demand), slash commands (procedures a person triggers), subagents (isolated side tasks), hooks (checks that always run), MCP servers (live external systems) and plugins (bundles of the rest). Choose by job, then measure context cost.

A colleague asks the agent to open a PR “the way we do it here”. The agent has three sources of instructions, two GitHub integrations and a forgotten hook, and it follows the wrong one. This page helps developers and tech leads decide where a rule or tool belongs.

Which mechanism does this job?

Mechanism	Enters context	Reach for it when	Not for	Canonical page
AGENTS.md (CLAUDE.md, Cursor Rules)	Every session, in full	build and test commands, rules for every task	long procedures, API docs	Repository context files
Skill (SKILL.md folder)	Name and description always, body when it fires	a repeatable procedure: TDD, a release checklist	credentials, live data	Agent skills
Slash command (in Claude Code, a skill with disable-model-invocation: true)	Only when a person types it	steps that must never start alone: a deploy, a release cut	anything the agent should pick up unprompted	Custom commands and skills
Subagent	Its own window; only a summary returns	noisy searches, a second-opinion review	work that needs the main thread’s context	Custom subagents
Hook	Nothing, unless it returns output	rules that must hold every time: format on edit, block a secret	advice the model may weigh	Hooks
MCP server	Tool schemas plus every result	reading or changing external state: PRs, tickets, a browser	conventions and style	MCP servers
Plugin (a versioned bundle of the above)	Whatever it bundles	a vendor’s skills and server together, or a team’s standard setup	a rule one line in AGENTS.md covers	Plugins

The test: a skill tells the agent how, an MCP server lets it act, and a hook makes sure it did. A rule the agent may ignore belongs in AGENTS.md or a skill; a rule it must never break belongs in a hook or in CI.

Note

Claude Code reads CLAUDE.md and falls back to AGENTS.md (v2.1.277 and later, latest channel). Codex skips AGENTS.md in an untrusted project (0.150.0 and later). In Claude Code, custom commands are skills, so Claude can invoke one on its own unless its frontmatter sets disable-model-invocation: true.

Where each layer lives in the ecosystem

Skills Practice skills, vendor skills by stack, installing and securing them.

MCP servers GitHub, Context7, Playwright and servers by job, plus token cost.

Plugins Install commands per agent, top plugins, team marketplaces.

Frameworks Superpowers, Spec Kit, OpenSpec, BMAD and GSD, shipped as plugins.

Agent tools herdr, desktop environments, sandboxes, model gateways.

Observability and quality Telemetry, cost, review bots, security gates, evals.

Models Current model names, prices and defaults.

What changed in 2026

• Plugins became the default install route. On 2026-09-26, claude-plugins-official listed 314 plugins, openai-curated 65 and cursor/plugins 94.
• Context cost became the selection criterion. claude plugin details (Claude Code 2.1.283, 2026-09-26): about 838 tokens per session for superpowers, 41,515 for Everything Claude Code.
• Churn is high. GSD is now @opengsd/gsd-core; openai/skills gave way to openai/plugins.

Equip one repo from zero

1. Record the baseline: /context in Claude Code. Codex and Cursor have no verified equivalent, so count servers with codex mcp list and skills with npx skills list. claude plugin details <name> prints any plugin’s projected token cost, even for skills you install elsewhere.

2. Install one discipline pack. Matt Pocock’s skills cover grilling, specs, tickets, TDD and review.

   Claude Code

   claude plugin install mattpocock-skills@claude-plugins-official --scope project
   claude plugin details mattpocock-skills

   Codex

   npx skills add mattpocock/skills --skill grill-me grilling setup-matt-pocock-skills -a codex

   Cursor

   npx skills add mattpocock/skills --skill grill-me grilling setup-matt-pocock-skills -a cursor

   grill-me is a one-line delegate that fails without grilling. Write --skill grill-me, never --skill=grill-me, which skills 1.7.0 drops silently. Plugin skills are namespaced: /mattpocock-skills:grill-me in Claude Code; an npx skills copy is /grill-me in Cursor and $grill-me in Codex.

3. Add GitHub and Context7. Read a fine-grained, single-repository token from a secret store (macOS Keychain here), never paste it. The GitHub MCP page lists minimum scopes and the /readonly URL.

   Claude Code

   export GITHUB_PERSONAL_ACCESS_TOKEN="$(security find-generic-password -s gh-mcp -w)"
   claude plugin install github@claude-plugins-official --scope project
   claude mcp add --transport http --scope project context7 https://mcp.context7.com/mcp

   Codex

   export GITHUB_PAT_TOKEN="$(security find-generic-password -s gh-mcp -w)"
   codex mcp add github --url https://api.githubcopilot.com/mcp/ --bearer-token-env-var GITHUB_PAT_TOKEN
   codex mcp add context7 --url https://mcp.context7.com/mcp

   Cursor

   Add Context7 to .cursor/mcp.json:

   { "mcpServers": { "context7": { "url": "https://mcp.context7.com/mcp" } } }

   For GitHub in Cursor, follow the token setup on the GitHub MCP page.

4. Measure again in a new session. If the cost outweighs the value, remove something.

5. Commit .claude/settings.json, .mcp.json and .cursor/mcp.json so the team gets the same setup through review. codex mcp add writes to your user config, so each Codex user repeats step 3. Each teammate still exports their own GITHUB_PERSONAL_ACCESS_TOKEN (the plugin reads exactly that name); tokens never go into the committed files.

Copy-paste prompt: audit this repo's agent setup

List every source of instructions and tools this session has: AGENTS.md or CLAUDE.md files, installed skills and plugins, subagents, hooks and connected MCP servers. For each, give its location and what it costs in context. Then flag: (1) any rule that appears in two places, (2) any skill installed twice (plugin and npx skills), (3) any MCP server that duplicates a CLI already on PATH, such as gh, (4) any "must" rule that lives only in prose and should be a hook or a CI check. Propose removals first, additions last. Do not change any file.

How you know the setup works without reading every line

• Context budget: the step 1 and step 4 numbers, recorded in the PR; see reducing MCP token cost.
• Deterministic gates: tests, type checks and a Gitleaks pre-commit hook.
• A second model: codex review or a review bot; see agent PR review.
• Sign-off: the tech lead reviews MCP and plugin changes as dependencies.

When a layered setup goes wrong

Symptom	Cause	Recovery
Every skill appears twice	Installed as a plugin and with npx skills	Keep one route
The agent ignores a “must” rule	The rule lives only in prose	Move it into a hook or a CI check
An MCP result asks the agent to do something odd	Prompt injection through fetched content	Use read-only scopes; see MCP security

Where to go next from the ecosystem overview

• Before this: Git worktrees for parallel agents, the previous step in the developer track.
• Plugins and marketplaces to pin versions per agent.
• Agentic development frameworks compared to choose one discipline pack; discipline packs walks one feature from grilling to verification.
• Model routing by evidence, the next step in the developer track.