Compare
A context system is not an alternative to skills and plugins. It uses them.
Where a context system sits next to standing rules, skills, plugins, subagents and MCP, in Claude Code and in Codex, using each tool's own words rather than a mapping invented to make the table look tidy.
If you have heard of skills and plugins and are trying to work out what this new noun is supposed to be, this is the page. The short answer is that skills and plugins are mechanisms a tool gives you, and a context system is what you build with them, plus two things neither tool provides.
The mechanisms
Five things people mix up, and what each one actually is.
Standing rules
A file the agent reads before it does anything, every session, without being asked. It is the only one of these that is always in context, which is why it has to stay short and why anything with a date on it does not belong in it.
Skills
A folder with a SKILL.md in it, holding a procedure the agent loads only
when it is relevant. Both tools work the same way here: the agent starts with just
the name, the description and the path, and reads the body only if it decides to use
it. That is what makes a system able to be thorough without being heavy.
Plugins
A distribution unit. It is how a skill, and in Claude Code a set of agents, hooks and MCP servers with it, gets installed into somebody else's setup. A plugin answers the question of how this reaches a colleague, not the question of what it knows.
Subagents
A delegated agent with its own instructions and often its own tools, started to handle one piece of work. Useful when you want a second opinion produced by something that has not seen the first one.
MCP servers
A protocol for handing the agent tools and data from outside itself. It extends what the agent can reach. It has nothing to say about what the agent should do with it.
Two vocabularies
They have converged more than the noise suggests.
The expectation going in was that these would be loose analogies. Mostly they are not. Skills in particular are close to identical in both tools, down to the filename and the frontmatter, and the real porting work is directory names.
| Concept | Claude Code | Codex | When it loads |
|---|---|---|---|
| Standing rules | CLAUDE.md |
AGENTS.md, or AGENTS.override.md where present. Files are concatenated from the repo root down to your working directory, and the total is capped by project_doc_max_bytes, 32 KiB by default |
Every session, before work starts |
| Skills | .claude/skills/<name>/SKILL.md in a project, ~/.claude/skills/ for personal |
.agents/skills/ in a repo, scanned from your working directory up to the repo root, and ~/.agents/skills/ globally |
Name, description and path up front; the body only when the agent decides to use it |
| Skill format | SKILL.md with YAML frontmatter carrying name and description, plus optional supporting files |
SKILL.md with YAML frontmatter carrying name and description, plus optional scripts, references and assets |
The same idea in both, and this is the strongest genuine equivalence on this page |
| Slash commands | Merged into skills. .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both produce /deploy, and existing command files keep working |
Codex documents custom prompts. Not read in detail, so no equivalence is claimed here | On invocation |
| Plugins | .claude-plugin/plugin.json, installed from a marketplace, able to bundle skills, agents, hooks and MCP servers. Plugin skills are namespaced plugin-name:skill-name |
Documented as the way skills are packaged and shared. Whether a Codex plugin can also carry subagents and MCP servers was not verified | On install |
| Subagents | .claude/agents/ |
TOML files in .codex/agents/ or ~/.codex/agents/, each requiring name, description and developer_instructions |
When work is delegated to one |
| MCP servers | Configured per project or per user | Configured in config.toml, at ~/.codex/config.toml or, for trusted projects, .codex/config.toml |
At startup |
| Hooks | Configured in settings | Documented, and not read in detail for this page | On the event they are bound to |
The one difference worth planning around is format rather than vocabulary: Claude Code defines a subagent in Markdown with frontmatter, Codex defines one in TOML with three required fields. Skills and standing rules port with a rename. Subagents get rewritten.
The answer
Three of the five parts are not tool features at all. They are just files.
This is the part that makes the category legible. Line the five parts of a context system up against what the tools actually provide, and most of it turns out not to be a tool question.
| Part | What the tool provides | What the context system supplies |
|---|---|---|
| Standing rules | The loading mechanism, and a filename | The actual rules for one domain, kept free of anything with a date on it |
| Skills | Progressive disclosure, so the body costs nothing until it is used | The phases, and the method inside each one |
| Knowledge base | Nothing. Neither tool has this concept | Files the rules instruct the agent to cite instead of recalling |
| Scripts | Nothing beyond the ability to run them | Committed code that produces every deliverable, so nothing is hand exported |
| Tests | Nothing. Neither tool has an opinion about whether your rules still hold | Tests that fail when a rule is broken, which is the only thing that stops rules rotting |
Which is why the interesting engineering is not in the tool
A context system is a repo. It uses the standing rules file and the skills mechanism your agent already has, and then it adds the knowledge base, the scripts and the tests, none of which either product provides or has any view about.
That is also why these systems are not really locked to one agent. The rules file gets renamed, the skills directory moves, and the three parts that carry most of the value do not move at all, because they were only ever files in a repo.
What was checked, and what was not
Sources, and the gaps in them.
Everything in the table above was read from current documentation on 2026-08-20 rather than recalled. Both products move quickly, so treat this page as dated and check the source when it matters.
[not verified] Codex hooks and Codex custom prompts were not read in detail, and no equivalence is claimed for either. Whether a Codex plugin can bundle subagents and MCP servers the way a Claude Code plugin can was not confirmed.
Sources: Claude Code, skills, Codex, customization overview, Codex, subagents, Codex, build skills, Codex, AGENTS.md.
Claude Code and Codex are the products of Anthropic and OpenAI respectively, named here because that is what they are called.