Codex vs Gemini CLI: which do top repos set up for?
Codex averages higher: 70.5 for Codex and 63.5 for Gemini CLI, across 402 repos. Codex scores higher on 397, Gemini CLI on 5, and 0 tie.
Scores measure how well each repo is set up for an agent: instruction files, tests, CI and the other checks, weighted by what each agent's docs say it uses, not which agent writes better code.
Which files does each one read?
- Reads from the repo
- Codex:
AGENTS.mdAGENTS.override.md.agents/skills/ - Gemini CLI:
GEMINI.md.agents/skills/ - Reads AGENTS.md
- Codex: Always
- Gemini CLI: Only when configured to
- Reads CLAUDE.md
- Codex: No
- Gemini CLI: No
- From its docs
- Codex: Reads AGENTS.md before any work, one per directory from the repo root down to the working directory, up to 32 KiB combined; AGENTS.override.md replaces the AGENTS.md beside it.
- Gemini CLI: Reads GEMINI.md from the global file, the workspace and its parent directories, and one per subdirectory once a tool reaches it. AGENTS.md loads once listed in the
context.fileNamesetting.
AGENTS.md and CLAUDE.md mean the file at the repo root; a repo with both counts under AGENTS.md, and empty files don't count. Every agent's files.
Which checks does each weigh more?
| Check | Codex | Gemini CLI |
|---|---|---|
| GEMINI.md | 0% | 11% |
| AGENTS.md / CLAUDE.md | 11% | 8% |
| CI configuration | 9% | 7% |
| Reproducible dev env | 9% | 8% |
| Type configuration | 9% | 10% |
| Manageable size | 6% | 5% |
Each figure is the check's share of the agent's total weight, from the docs linked below. A check few repos pass, such as an agent's own file, lowers that agent's scores wherever it carries weight. The methodology lists every weight.
Which repos are set up better for Codex than Gemini CLI?
- hiyouga/LlamaFactory87.2 vs 76.0 for Gemini CLI
- langflow-ai/langflow83.0 vs 72.1 for Gemini CLI
- gitlab-org/gitlab85.2 vs 74.3 for Gemini CLI
- scikit-learn/scikit-learn81.8 vs 71.0 for Gemini CLI
- cline/cline81.4 vs 70.7 for Gemini CLI
Which repos are set up better for Gemini CLI than Codex?
- google-gemini/gemini-cli82.4 vs 77.3 for Codex
- apache/seatunnel71.4 vs 68.4 for Codex
- rustdesk/rustdesk74.1 vs 71.5 for Codex
- apache/superset91.3 vs 89.9 for Codex
- discourse/discourse93.4 vs 92.3 for Codex
Which one leads in each language?
| Language | Repos | Codex | Gemini CLI |
|---|---|---|---|
| TypeScript | 83 | 75.7 | 67.6 |
| Python | 56 | 74.1 | 66.9 |
| Rust | 45 | 74.0 | 67.5 |
| Go | 40 | 74.6 | 67.6 |
| Java | 28 | 65.5 | 58.8 |
| C++ | 21 | 62.4 | 55.5 |
| JavaScript | 15 | 77.1 | 69.6 |
What if I use both?
- Write the shared instructions once, in AGENTS.md at the repo root. What to put in it.
- Codex: reads AGENTS.md before any work, up to 32 KiB combined across directories.
- Gemini CLI: list AGENTS.md in the
context.fileNamesetting, or keep a GEMINI.md.
AGENTS.md vs CLAUDE.md · Which files each agent reads · Make a repo agent-ready
From the docs: learn.chatgpt.com/agents-md (new tab) · learn.chatgpt.com/build-skills (new tab) · geminicli.com/gemini-md (new tab) · geminicli.com/skills (new tab)
Boards: best set up for Codex · best set up for Gemini CLI
More comparisons: Claude Code vs Cursor · Claude Code vs Codex · Gemini CLI vs Claude Code · Cursor vs Codex · Aider vs Claude Code · OpenHands vs Claude Code · Pi vs Claude Code · Kimi Code CLI vs Claude Code · Devin vs Claude Code