Claude Code vs. GitHub Copilot vs. OpenAI Codex

This documentation was generated with the assistance of AI. Please report any inaccuracies.
This is a snapshot of a fast-moving area — every provider here has changed this surface within the last year. Treat the specific file paths and command names below as a starting point to verify against each provider’s current documentation, not a permanent contract.

The four artifacts this catalog is built around (CLAUDE.md, skills, agents, plugins) aren’t unique to Claude Code — GitHub Copilot and OpenAI Codex have converged on broadly similar concepts, sometimes under the same file name, sometimes not. This page maps one onto the other so a decision made here (what belongs in memory versus an on-demand skill, how far to isolate a sub-agent) transfers if the same team also uses another assistant.

At a glance

Concept Claude Code GitHub Copilot OpenAI Codex

Persistent, always-loaded project memory

CLAUDE.md (repo root, user-level ~/.claude/CLAUDE.md, nested subdirectory files, enterprise policy — The CLAUDE.md file)

.github/copilot-instructions.md — broadest support across Chat/IDE/coding agent/review. Increasingly supplemented by the cross-tool AGENTS.md convention, though AGENTS.md support is currently narrower (mainly the cloud coding agent; VS Code support is experimental/opt-in).

AGENTS.md — Codex’s primary mechanism. Read hierarchically: a global ~/.codex/AGENTS.md, plus one file per directory from the project root down to the working directory, all concatenated.

Scoped / path-specific instructions

Nested CLAUDE.md files load only when that subtree is touched; .claude/rules/ glob-scoped rule files.

.github/instructions/*.instructions.md, with an applyTo: glob in its frontmatter. Honored broadly in IDEs; on github.com itself only by the coding agent and code review, not plain Chat there.

An AGENTS.override.md in a given directory replaces (rather than adds to) the plain AGENTS.md at that same level — an explicit escape hatch rather than a glob-matched rule file.

On-demand, explicitly invoked playbook (this catalog’s "skill")

.claude/skills/<name>/SKILL.md, invoked as /<name> — Creating and updating skills and agents

.github/prompts/*.prompt.md in VS Code/Visual Studio, filename becomes the /name command. Not yet recognized by the Copilot CLI as of this writing.

Codex "skills" — packaged instructions (+ resources/scripts), invoked explicitly or matched implicitly to a task. Superseding the now-deprecated ~/.codex/prompts/*.md custom-prompt mechanism.

Custom sub-agent / persona with its own tools

.claude/agents/<name>.md — Agents

Custom agents: an .agent.md file under .github/agents/ (or an org-wide .github/agents repository under the organization account itself), auto-delegated to as an isolated sub-agent. .github/chatmodes/.chatmode.md covers a lighter-weight persona/system-prompt switch without full sub-agent isolation.

Built-in default/worker/explorer agents, plus custom agents as TOML files under ~/.codex/agents/ (personal) or .codex/agents/ (project).

Packaging and distributing across repositories

Plugins — a manifest bundling skills/agents/commands/hooks/MCP config, installed via /plugin install from a public or private marketplace (Plugins)

No direct equivalent today — GitHub deprecated its GitHub-App-based Copilot Extensions marketplace (Nov 2025) in favor of plain MCP. The closest substitutes are org-wide custom instructions (GA since Apr 2026) and community-curated instruction/prompt repositories, neither of which is an installable versioned package.

Codex plugins (introduced around v0.117.0) bundle skills, tool integrations, MCP config, and hooks, distributed via marketplace.json sources — a genuine analog to Claude Code plugins, though without one centralized official marketplace yet.

MCP server configuration

.mcp.json / plugin .mcp.json / settings

.vscode/mcp.json (or a user-profile file) keyed by "servers"; the Copilot CLI has its own mcp-config.json — configuration is per-surface, not unified.

~/.codex/config.toml, under [mcp_servers.<name>] — stdio or HTTP transport.

Where the underlying models genuinely differ

  • Codex is the closest structural match to Claude Code. AGENTS.md ~ CLAUDE.md, Codex skills ~ Claude Code skills, Codex sub-agents ~ Claude Code agents, and Codex plugins ~ Claude Code plugins — the two tools converged on nearly the same four-artifact shape independently. If a team already writes clean, step structured skills for Claude Code (Authoring effective instructions), most of that same content ports to Codex with only mechanical translation.

  • Copilot’s model is scope-matched instructions layered across multiple product surfaces (Chat/IDE, the cloud coding agent, code review) rather than one unified agent runtime, and — notably — each surface currently honors a different subset of the instruction files above. Porting a Claude Code skill/CLAUDE.md to Copilot means checking which surface the team actually uses it from, not just picking the file with the closest name.

  • Only Claude Code and (more recently) Codex treat packaging/distribution as a first-class, versioned concept (a plugin manifest, a marketplace, an install command). Copilot dropped its nearest equivalent rather than converging toward one, leaving org-wide sharing there closer to "copy the file" (this catalog’s own Option 1 — Packaging and distributing skills, agents, and CLAUDE.md) than to Option 2.

  • The AGENTS.md convention is a genuine point of convergence worth watching: Codex treats it as canonical, and Copilot has started reading it as an alternative to copilot-instructions.md on some surfaces. A repository that wants one memory file to mean something to more than one assistant should keep an eye on whether that convention keeps spreading, rather than assuming CLAUDE.md, AGENTS.md, and copilot-instructions.md will stay three separate files indefinitely.

See also