Bash (Anthropic — no standalone skill in anthropics/skills)

tl;dr: there is no SKILL.md you can install for bash in anthropics/skills. The repo’s 19 skill folders are all document, creative, dev, or enterprise — no shell. Bash runs as a native tool in Claude Code or as part of the computer use toolset.

Status: not in the skill repo

anthropics/skills (175.5k stars, 20.8k forks) is the canonical Agent Skills repo, but its skills/ directory contains only the following as of this fetch:

academy-guide, algorithmic-art, brand-guidelines, canvas-design, claude-api, discernment-nudge, doc-coauthoring, docx, frontend-design, internal-comms, mcp-builder, pdf, pptx, skill-creator, slack-gif-creator, theme-factory, web-artifacts-builder, webapp-testing, xlsx

(Bolded: the four document skills that ship with proprietary LICENSE.txt.)

So bash lives in two non-skill surfaces, both Anthropic-maintained:

If you were looking for a SKILL.md you can install via /plugin install, it doesn’t exist. Install the broader plugin (document-skills or example-skills) or use Claude Code’s native tool.

Claude Code native Bash tool

Claude Code ships a persistent-shell Bash tool with a per-command timeout, working directory persistence, and explicit permission rules. The tool runs commands in a sandboxed subprocess and returns stdout + stderr + exit code.

# Install Claude Code
npm install -g @anthropic-ai/claude-code

# Run a session (Bash is one of the default tools)
claude

Current: Claude Code 2.1.267 (latest from the upstream CHANGELOG.md; substantive changes in the last few weeks — 2.1.265 added a 1 GB cap on tool results saved to disk and --plugin-dir for plugin folders; 2.1.267 added maxEffortLevel setting and --system-prompt-snapshot off for prompt iteration, plus many prompt-cache and resume fixes). Repo: anthropics/claude-code at 145k stars.

Safety: outside the demo container, the Bash tool executes with your user permissions — treat it as arbitrary code execution. Configure permissions in ~/.claude/settings.json to allow/deny specific patterns. Recent versions tightened the bash permission checks (2.1.259 covers cd DIR && cat FILE, < file redirects, grep -r, cp -r; 2.1.260 covers Bash(ls) x junk rules, rm -rf on positional args, and Read() deny rules covering file arguments).

Computer use bash (via the demo)

The computer-use-demo no longer ships a separate bash tool. With computer_toolset_20260801 (out of beta on August 19, 2026), the computer tool is a toolset: one entry in tools[] with member tools per action. tool_use.name is the member (left_click, screenshot, type, …), no action discriminator, and toolset_name: "computer". To run a shell command in computer use, you drive a terminal inside the desktop.

docker run \
  -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY \
  -v $HOME/.anthropic:/home/computeruse/.anthropic \
  -p 5900:5900 -p 8501:8501 -p 6080:6080 -p 8080:8080 \
  -it ghcr.io/anthropics/anthropic-quickstarts:computer-use-demo-latest

What to install instead

• Claude Code for local agentic shell work — /plugin install example-skills@anthropic-agent-skills adds example skills on top.
• Computer use demo for desktop/screenshot-driven shell work — see the README for screen-size and security caveats.
• Code execution tool (code_execution_20250825) on the API if you want a sandboxed REPL with file output via the Files API, without driving a desktop. Available on Claude Opus 5 / Sonnet 5 / Haiku 4.5 (and the Fable/Mythos tiers). Same tool ID is required for any skill you attach via container.skills.

Gotchas

• Don’t grep anthropics/skills for “bash” — the SKILL.md doesn’t exist; you’ll waste a fetch.
• Computer use demo screen sizes are not free. Pass a computer-use-demo --width/--height that matches your X server or VNC client; mismatched sizes degrade the screenshot classifier.
• API key goes through ANTHROPIC_API_KEY in the demo container. Do not commit the token; the demo’s VNC ports (5900, 6080) are exposed by default.
• Bash tool is sandboxed via Docker in the demo but uses your user’s permissions in Claude Code. Treat it as arbitrary code execution.
• Claude Code changelog is dense. 2.1.267 (current) adds maxEffortLevel (caps effort per provider) and --system-prompt-snapshot off for prompt iteration. Recent substantial additions: 1 GB cap on tool results, --plugin-dir, --append-subagent-system-prompt-file, /skill-doctor, /advisor text form, /reload-plugins, bashOutputMaxChars/taskOutputMaxChars.

Related

• Anthropic PDF (official skill) — same skills shelf, also a routing-only skill
• frontend-design (Anthropic, UI Skills) — same Anthropic skills shelf
• UI Skills catalog (Anthropic) — broader Anthropic UI skills surface
• Anthropic Claude — what Anthropic publishes — cross-shelf context