Library for Claude Code and Codex

Reviews the diff, a PR or a branch for correctness bugs. --fix applies the fixes and --comment posts them on the PR.

Four parallel agents look for reuse, simplification, efficiency and abstraction level, then apply fixes. It does not hunt bugs.

Scans the branch's changes against origin for vulnerabilities.

Repeats a prompt or command while the session stays open; without an interval Claude paces itself (1 min to 1 h). Good for watching a deploy or CI.

Creates, lists and runs scheduled cloud routines (needs a claude.ai login).

Large-scale changes: 5 to 30 independent units, one subagent per unit in its own worktree.

Fan-out web searches, source cross-checking and a cited report.

/run launches the project's app so you can see a change working; /verify checks that the change does what it should.

Generates a starter CLAUDE.md with what Claude needs to know about the project.

Claude edits the right settings.json for you: permissions, environment variables and hooks.

Scans your transcripts and proposes a read-only allowlist so you stop getting interrupted.

Loads the Claude API reference (models, caching, tool use, agents); it triggers by itself when you import the Anthropic SDK.

Creates and improves your own skills with the right structure.

Create and edit Word documents, PDFs, slide decks and spreadsheets.

Guide to building high-quality MCP servers.

Engineering pack: structured debugging, review, architecture decisions, incident response and deploy checklist.

Answers questions about Codex models, pricing, skills, settings and automations from the official docs.

Create Codex skills and install them from a curated list or a GitHub repo.

Read-only, defect-first review of a specific change; returns every actionable finding.

Generates or edits raster images (photos, illustrations, textures, mockups) when a task needs them.

Document plugins: Word/Docs with render-and-verify, PDF (including forms), PowerPoint/Slides and Excel/CSV.

Drive Windows apps and the browser from Codex.

Builds visualisations and interactive tools right in the conversation.

“Superpowers” pack: explore requirements before coding, write the plan, and work test-first.

Root-cause debugging before proposing fixes, and demanding evidence (passing tests) before calling work done.

Run plans with subagents, parallelise independent tasks, and isolate work in Git worktrees.

codex-security plugin: phased security scan (threat model, discovery, validation) and verified fixes.

plugin-eval plugin: scores a skill or plugin, measures real token use and turns findings into an improvement brief.

build-web-apps plugin: frontend apps from scratch, React/Next.js best practices (Vercel) and shadcn components.

cloudflare plugin: Workers and wrangler CLI, best practices and web performance (Core Web Vitals).

New conversation with an empty context (aliases /reset, /new).

Summarises the conversation to free context without losing the thread.

Shows how much context is in use, with suggestions.

Resume a conversation, branch it to try another path, or copy it to a background session.

Rewinds the conversation and/or the code (aliases /checkpoint, /undo).

Pick the model, effort level (low to max) and fast mode.

Session cost, plan limits and stats (aliases /cost, /stats).

Manage allow/ask/deny rules, directories and auto mode.

Edit CLAUDE.md files and automatic memory.

Manage MCP servers and their authentication.

Manage plugins, list skills (with context cost) and view hooks.

/goal: Claude keeps going until a condition is met. /plan: enters plan mode.

Side question that does not enter the conversation history.

Moves the session to the background and shows running work (shells and subagents).

Grants access to another directory or changes the working directory.

Settings, status line and terminal setup (Shift+Enter, etc.).

Review the project's changes; export the conversation as text.

New chat in the same session / clear the screen and start over.

Resume a saved session, fork it, or open an ephemeral side chat (/btw).

Summarises the visible chat to free tokens.

Pick the model and reasoning effort; toggle the Fast tier when offered.

Plan mode, and a task goal you can pause and resume.

Approval and sandbox presets: Auto, Read Only, Full Access.

Review uncommitted work, a commit or the base branch; show the Git diff.

Generates a starter AGENTS.md in the current directory.

Session configuration, token use and limits.

See MCP tools, connectors, plugins and skills; skills are also invoked with $name.

Manage lifecycle hooks and memory usage.

Attach a file to the chat; import settings and chats from Claude Code or Cursor.

See and stop background terminals.

Remap shortcuts, enable Vim mode and configure the status line.

Continues the current session in the ChatGPT desktop app.

Cycles permission modes (normal, accept edits, plan…).

Interrupts without losing work; twice on an empty prompt opens the rewind menu.

Toggles the transcript: tools, timestamps and model per message.

Reverse search in prompt history.

Opens the prompt in your text editor (handy for long prompts).

Sends Bash commands and agents to the background.

Shows or hides Claude's task list.

Switch model, toggle extended thinking and fast mode without losing your draft.

Pastes an image from the clipboard into the prompt.

At the start: ! runs a shell command, @ mentions a file, / opens commands and skills.

New line in the prompt without sending.

Stops all background subagents (twice within 3 s).

Lists every shortcut in the Claude desktop app.

Desktop app: new session, close session, next session.

Desktop app: diff pane, browser pane and terminal.

Toggles focus between editor and Claude; Alt+K inserts an @-mention of the file and selection.

@ finds project files; a leading ! runs a local command under the current approval and sandbox.

While Codex works: Tab queues your instruction for the next turn; Enter injects it into the current turn.

On an empty prompt: edit your previous message and fork from there.

Search history; copy the last output; clear the view without losing the chat.

Shows shortcut help in the composer.

The app's command palette; Ctrl+/ lists every shortcut.

Switches between Chat, Work and Codex modes.

New chat, side chat and next chat needing attention.

Model picker and recent models.

Bottom panel, integrated terminal and file search.

Review tab and new browser tab.

Project instruction file: Claude Code reads CLAUDE.md and Codex reads AGENTS.md. Both are generated with /init.

Configuration: ~/.claude/settings.json (Claude Code) and ~/.codex/config.toml (Codex).

In both, a skill is a folder with a SKILL.md (name + description + instructions). The agent picks it by its description, so a good description is half the job.

Connects external tools: “claude mcp add …” in Claude Code and “codex mcp add …” in Codex; /mcp shows status.

Claude Code: permission modes (Shift+Tab) and allow/deny rules. Codex: read-only / workspace-write / danger-full-access sandbox plus an approval policy (on-request, never).

Non-interactive mode for scripts and CI: “claude -p "task"” and “codex exec "task"” (with --json and --output-schema in Codex).

Scripts the tool itself runs on events (before/after a tool, on finish…). The right way to say “whenever X, do Y”: model memory is not enough.

Parallel work without stepping on files: each agent in its own Git worktree. Claude Code has /batch and subagents; Codex has /worktree, /side and subagent skills.

Claude Code: /schedule (cloud routines) and /loop (while the session is open). Codex: Automations (“Scheduled”) in the desktop app.