Codex vs Claude Code for Presentations: Which Coding Agent Makes Better Slides? (2026)

Both OpenAI Codex and Claude Code produce real PowerPoint files in 2026, and they do it the same three ways: an official first-party skill, community

packages, or a hosted generation service. The real differences are narrower than the marketing suggests: Codex ships a Slides skill with an installer ecosystem (

$skill-installer

); Claude Code ships Anthropic's pptx skill, which is the only first-party option that can edit an existing PPTX in place. For new decks at production quality, both agents do their best work when they delegate rendering — which is why the hosted path behaves almost identically in either.

This comparison covers only the presentation use case. For general coding comparisons, look elsewhere; for which Claude model to pick inside Claude Code, see which Claude model is best for presentations.

The three paths, in both agents

The community layer is effectively shared:

is a portable convention, so codex-ppt-skill, the nano-banana image skills, and the 2Slides skill run in both agents. The eight-skill comparison applies to either.

Where they genuinely differ

1. Editing existing decks: Claude Code wins

Anthropic's pptx skill unpacks and rewrites slide XML, which means "fix the typo on slide 7 of this client deck and update the Q3 numbers" works on a file the agent didn't create. Codex's official skill is generation-first; editing an existing deck means regenerating or hand-writing python-pptx code. If your workflow is maintaining decks rather than producing them, this single capability decides the comparison.

2. Skill distribution: Codex wins on ergonomics

plus a restart is the smoothest install flow in any agent today, and OpenAI's skills repo formalized the packaging. Claude Code's folder-drop convention works fine but is more manual. If your team shares custom skills widely, Codex's tooling is ahead.

3. Code-composed output quality: a coin flip, honestly

Both official skills write PptxGenJS and check their own rendering. Output quality tracks the model's layout decisions that session more than the agent brand — both produce clean-but-engineered decks, and both fall short of designed templates on visual polish. We do not publish head-to-head scores here because single-run anecdotes don't generalize; run your own brief through both before standardizing.

4. Beyond PPTX: a tie via the hosted path

Neither official skill does themed template libraries, reference-image style matching, multi-language output, or narration. In both agents, those arrive the same way — a hosted pipeline behind a skill or MCP server. With the 2Slides skill installed, the Codex-vs-Claude-Code question stops mattering for output: same templates, same 19+ languages, same narration and per-page exports, same credits (endpoint reference).

Setup, side by side

Codex — official skill needs nothing; hosted skill:

export SLIDES_2SLIDES_API_KEY="sk-..."
$skill-installer https://2slides.com/skill.md
# or MCP in ~/.codex/config.toml:
# [mcp_servers.2slides]
# command = "npx"
# args = ["2slides-mcp"]

Claude Code — official pptx skill is available out of the box in current builds; hosted path:

export SLIDES_2SLIDES_API_KEY="sk-..."
# skill: clone github.com/2slides/slides-generation-2slides-skills into your skills dir
# or MCP:
claude mcp add 2slides -- npx 2slides-mcp

Full walkthroughs: Codex + skills, Codex + MCP, Claude Code + MCP.

Which should you use?

• You maintain existing client decks → Claude Code; XML-level editing is unmatched.
• Your team distributes custom skills at scale → Codex;
  $skill-installer
  ergonomics compound.
• You already live in one agent → stay there; add a hosted skill and the output gap closes to zero.
• You're choosing an agent primarily to make presentations → don't. Choose the agent for your coding workflow and give it a slide engine; the engine, not the agent, determines deck quality.

FAQ

Can Codex and Claude Code both output real editable PPTX?

Yes. Both official skills compose native PPTX with real text boxes (PptxGenJS under the hood), and both verify rendering before returning the file. Image-composed community skills are the exception — those bake text into pixels in either agent.

Which agent makes better-looking slides?

With official skills: comparable, and both plainer than designed templates — quality varies run to run with the model's layout code. With a hosted skill, output is template-locked and identical across both agents, which is the point.

Do Codex skills work in Claude Code and vice versa?

Community

packages generally run in both (codex-ppt-skill, 2Slides skill, and most directory listings document multi-agent support). The two official skills are agent-locked: OpenAI's Slides skill to Codex, Anthropic's pptx skill to Claude.

Can either agent edit a PowerPoint file it didn't create?

Claude Code can, through Anthropic's pptx skill's XML rewriting. Codex has no equivalent first-party capability for in-place PPTX editing as of mid-2026.

What about Gemini CLI, OpenClaw, or Cursor?

The same pattern extends: all support MCP, and OpenClaw/Cursor read

. Setup guides: Gemini CLI, any MCP agent, and the skills vs MCP decision guide.