Claude Codex Settings
Battle-tested Claude Code, Codex CLI, and Cursor configs, plugins, hooks, and agents with Kimi, MiniMax, and GLM API support — fixing the failure modes of AI coding assistants.
README names hooks (guard.py, humanize.py, telemetry scripts) but their source is not in evidence, so least privilege, user confirmation, data flow, sensitive data handling, dependency security, external effects, and rollback cannot be substantiated; deducted for unverifiable hook behavior. Source attribution is partially supported: Karpathy quote, Wikipedia word-list source, anthropics/skills and openai skills credits with links.
The tool-support matrix is internally consistent (explicitly marks n/a per tool), earning 2 on self-consistency; dependency availability rests on assertions (signed-in CLIs, npx, external APIs) and failure messages are nowhere shown, hence deductions.
Audience and multi-tool coverage are clearly stated; capability boundaries are explicitly declared (e.g., intelligent-compact Claude-Code-only), which justifies 2. Deducted on trigger_precision because hook trigger conditions cannot be checked without hook source.
Structured README, per-platform install blocks, support matrix, stated platform limitations; full Apache-2.0 text present (score 3). Deducted: no CHANGELOG or versioning in evidence, no maintenance/contribution commitments, no FAQ or worked examples, and the README evidence is truncated at python-skills, indicating an architectural gap.
Each plugin has a distinct, plausible purpose (porting built-in capabilities across tools); marginal value is well argued. Deducted on cost_benefit: claims like lower token usage are qualitative only, and effectiveness cannot be verified without execution.
Descriptive claims map to in-repo paths (skills/hooks are all linked), and facts are mostly separated from inference; deducted for unverifiable claims ('battle-tested', traffic badges) and the inability to cross-check hook behavior against README descriptions.
- Static review only; no hook script was executed. The actual permissions and data flow of guard.py, humanize.py, and telemetry scripts are unverified — read their source before installing.
- Publisher is unverified by the FollowAgents registry; identity is unknown. Installing via marketplace implies trusting future pushes; prefer pinned installs.
- No CHANGELOG or version pinning mechanism is in evidence; hook behavior may change on any push.
- claude-telemetry-hooks emits session and rejection-categorization data; confirm its output destination matches your data policy before deploying.
- guard.py blocks commits until a /simplify run completes and could disrupt normal workflows; verify its skip mechanism before adoption.
What does this agent do, and when should you use it?
fcakyon/claude-codex-settings is a curated, battle-tested configuration collection for Claude Code, Claude Desktop, OpenAI Codex, and Cursor, shipping skills, commands, hooks, subagents, and MCP servers. It is distributed as a plugin marketplace: add it once (`claude plugin marketplace add fcakyon/claude-codex-settings`) and install individual plugins by name. Flagship plugins include simplify (four-angle code-review cleanups enforced by a pre-commit hook), humanize (a hook that blocks AI buzzwords on every write), fable-advisor and codex-advisor (cross-model second opinions), intelligent-compact (priority-preserving session summaries), claude-telemetry-hooks (usage and rejection telemetry), and large official skill bundles from Anthropic, OpenAI, Vercel, MongoDB, Supabase, Stripe, Cloudflare, and Hetzner. A shared CLAUDE.md guidance file can be symlinked as AGENTS.md and GEMINI.md so Codex, Cursor, Copilot, and Gemini CLI follow the same rules. It is Apache-2.0 licensed and listed in Awesome Claude Code.
The repository is not a standalone service; it acts as a plugin marketplace and config source for AI coding tools. Once installed: the simplify plugin's /simplify command reviews staged or committed diffs across four angles (reuse, simplification, efficiency, altitude) with parallel agents and applies cleanups, while its guard.py hook requires a completed /simplify run before each commit unless explicitly skipped. The humanize plugin's humanize.py hook scans every Write or Edit before it lands and blocks AI-writing markers — the em-dash, 53 stock words like 'leverage' and 'seamlessly', 16 openers/cliches, and pile-up words — each with a plain-word swap. fable-advisor and codex-advisor send the recent conversation to Claude Fable (via claude -p --model fable) or GPT (via codex exec) for a challenge-your-plan review. intelligent-compact's precompact_priorities.sh hook instructs the /compact summarizer to preserve file paths under investigation, confirmed root causes, open questions, and expensive subagent findings. claude-telemetry-hooks emits a sticky per-project chat_id (SessionStart hook) and categorizes tool-rejection reasons (UserPromptSubmit hook) for OpenTelemetry dashboards. Skill bundles (python-guidelines, react-best-practices, pdf/docx/xlsx/pptx, mongodb, supabase, stripe, cloudflare, hetzner, openobserve, etc.) ship as SKILL.md files, downloadable ZIPs, and npx skills add targets; browser automation uses vercel-labs' agent-browser CLI (snapshot + element refs, claimed 93% less context than Playwright MCP).
- Developers on large codebases who want automated simplification review enforced on every commit via the simplify plugin's guard hook
- Engineers writing docs or blog posts who want the humanize hook to automatically block words like 'leverage' or 'seamlessly' in every write
- Teams that want a second model to challenge a plan before committing, installing fable-advisor or codex-advisor for cross-model review
- Long-session Claude Code users who lose context to auto-compaction and want intelligent-compact to preserve file paths, root causes, and open questions
- Teams building Claude Code usage dashboards, combining claude-telemetry-hooks with openobserve-skills for session and rejection analytics
- Multi-tool users (Claude Code + Codex + Cursor + Gemini CLI) who want one shared behavior spec via CLAUDE.md symlinked as AGENTS.md and GEMINI.md
What are this agent's strengths and limitations?
- One plugin set spans four tools — Claude Code, Codex CLI, Cursor, and Gemini CLI — e.g. porting Claude Code's built-in /simplify to the three tools that don't ship it natively
- Enforcement happens at the tool layer via hooks, not prompts: humanize blocks every offending write, guard requires /simplify before commits, so the model can't ignore the rules
- Aggregates official skills from Anthropic, OpenAI, Vercel, MongoDB, Supabase, Stripe, Cloudflare, and Hetzner, patched for CLI workflows (e.g. livekit-skills BYOK support)
- agent-browser uses snapshot + element refs instead of full DOM dumps, claiming 93% less context usage than Playwright MCP
- Listed in Awesome Claude Code, with llms.txt and a Context7 MCP index for easy retrieval
- Coverage is uneven per tool: intelligent-compact is Claude Code only, humanize runs on Claude Code and Gemini only, and claude-telemetry-hooks doesn't support Codex
- Cross-model advisors add credential and CLI dependencies (codex-advisor needs a signed-in Codex CLI; fable-advisor needs an authenticated Claude Code CLI from other tools)
- Claude Code output styles don't exist in Codex, Cursor, or Gemini, so adhd-output-style degrades to a skill there
- Parts of the documentation are incomplete in the available source (INSTALL.md prerequisites, the full hetzner-skills table), so verify before adopting
- Multiple hooks are invasive to existing CI or commit workflows and need to be evaluated one by one
How do you install or deploy this agent?
Prerequisites are in the repo's INSTALL.md. For Claude Code:
bash
# Add marketplace (one time)
claude plugin marketplace add fcakyon/claude-codex-settings# Install any plugin by name
claude plugin install fable-advisor@claude-settingsCodex CLI: codex plugin marketplace add fcakyon/claude-codex-settings, then codex plugin add simplify@claude-settings. Gemini CLI: gemini extensions install --path ./plugins/simplify. Cursor: cursor-agent plugin marketplace add https://github.com/fcakyon/claude-codex-settings, then install from the /plugin marketplace. Skill bundles can also be bulk-installed with npx skills add <repo-subpath> --skill '*'. Cross-model advisor plugins require the corresponding CLI signed in (codex-advisor needs a logged-in Codex CLI).
How do you use this agent?
Per-tool entry points: in Claude Code desktop, select + beside the prompt, then Plugins; in the Codex desktop app open Plugins and select Install; in Cursor open Customize and select Install. Typical invocations: run /simplify in Claude Code, Codex, Cursor, or Gemini to review and clean up changes; say 'ask the codex advisor' to have codex-advisor challenge the current plan; run /claude-tools:sync-claude-md to pull the shared guidance into ~/.claude/CLAUDE.md, or place CLAUDE.md in a project and symlink it (ln -sfn CLAUDE.md AGENTS.md, ln -sfn CLAUDE.md GEMINI.md) for Codex, Cursor, Copilot, and Gemini. Hook-based plugins (humanize, guard, intelligent-compact) activate automatically on every Write/Edit, commit, or compaction — no manual invocation needed.
How does this agent compare with similar options?
The README explicitly contrasts this repo with Claude Code's built-in /simplify (ported here to Codex, Cursor, and Gemini) and with Playwright MCP (agent-browser claims 93% less context usage). It also repackages official upstream skill repos — anthropics/skills, openai/plugins, vercel-labs/agent-skills, and others — positioning itself as a unified, patched distribution layer for multi-tool setups.
FAQ
Do I have to install everything?
What does it cost and what credentials are required?
npm i -g agent-browser && agent-browser install, and web-performance-skills bundles the chrome-devtools MCP server with no API key needed.I use Cursor or Gemini CLI — will features be missing?
Will hooks disrupt my workflow?
How do multiple tools share one behavior spec?
/claude-tools:sync-claude-md to sync the repo's .claude/CLAUDE.md to ~/.claude/CLAUDE.md, or place it in a project and symlink it as AGENTS.md and GEMINI.md.