claude-code-essentials

A curated, tested combination of Claude Code skills, MCP servers, and CLI tools, installed with one script and explained with one README. Not an exhaustive list of everything that exists. Every tool here was picked over real alternatives, and the reasons are documented below, not just asserted.

Tested against itself: this repo's own components were exercised while building it, not just described. code-review-graph indexed this exact repository. rtk was verified compressing real command output (a raw pip list at 19,520 bytes came back at 9,395 bytes through it). The .mcp.json merge logic was tested against this project's actual existing config, including a real code-review-graph entry with a custom env field, confirming nothing gets overwritten. Where something didn't work as expected during that testing, for example rtk's automatic hook not firing in every environment, it's documented as a known limitation rather than left out.

Demo

Real output from an actual --dry-run, not staged. Trimmed to a representative excerpt; the full run covers all 9 tools.

Who this is for

Solo developers using Claude Code who want a working, opinionated setup today, without evaluating a dozen plugins themselves first. If you don't have /plugin access in your environment (some hosted or SDK-based Claude Code sessions don't expose it), everything here still installs, since none of it depends on that command.

Not a fit if: you're coordinating a team of agents (see ruflo in docs/COMPARISONS.md, a better shape for that case), or you specifically want the broadest possible token-saving coverage over a proven track record (see the Token-max combination below instead).

Trust and safety before you run this

• The script fetches from each tool's original upstream source at install time. Nothing is vendored or bundled; you're always getting the real project, not a copy.
• pipx isolates every Python-based tool in its own environment so nothing here touches your other projects' dependencies (see why below).
• No secrets, API keys, or credentials are requested or read anywhere in this script.
• Piping curl | bash runs a script you haven't read yet. Standard advice applies here too: open install.sh and read it first, or download it and run bash install.sh locally instead of piping directly.

Tested on

• Windows: Git Bash. PowerShell is not supported for install.sh; run it through Git Bash or WSL instead.
• WSL, macOS, and Linux: standard bash, should work as-is.
• Claude Code: verified in a standard claude CLI terminal session and in at least one hosted/SDK-based desktop session. The rtk automatic hook behaves differently between the two; see the note below.

Install

curl -fsSL https://raw.githubusercontent.com/rishindra-mateti-tech/claude-code-essentials/main/install.sh | bash

Run it from inside the project you want MCP servers registered for. Restart Claude Code afterward: skills, commands, and MCP servers load at session start, not mid-session.

Flags: --dry-run (show what would happen, install nothing), --uninstall (remove everything, see docs/UNINSTALL.md), --skip-rtk, --skip-gsd, --skip-superpowers, --skip-graphify, --skip-markitdown, --skip-context-mode, --skip-taste-skill. Full list in docs/INSTALL.md.

The script reports success and failure per component honestly. If a step fails, it's listed under "FAILED" at the end, not silently counted as installed.

Quick example

Before this stack, a typical session looks like:

> git status
[300+ lines of raw output flood the context]
> "what calls this function?"
[Claude reads 8 files manually to find out]

After:

> git status
[rtk compresses it to a summary before it reaches context]
> "what calls this function?"
[code-review-graph answers from its pre-built index, no file reads needed]
> [writing a new helper function]
[ponytail checks: does this need to exist, or is it one line of stdlib?]

Same conversation, same questions, fewer tokens spent getting there, and code that defaults to minimal instead of defaulting to written-from-scratch.

How it's structured

Two tiers, on purpose. Some tools fire automatically because they're cheap and low-risk. Everything else stays dormant until you name it, so nothing runs behind your back that you didn't ask for.

The always-on tools were picked specifically because each one has a real cost/benefit case for running unattended: ponytail only activates on code writes, taste-skill only activates on frontend/UI writes, code-review-graph only responds to queries you make anyway, rtk only touches bash output. Nothing in the always-on tier watches or logs anything beyond what a normal session already produces.

What gets installed, and why

For the longer version of every row below, with metrics and the full reasoning against each alternative, see docs/TOOLS.md and docs/COMPARISONS.md.

Tool	Role	Chosen over	Why
ponytail (4 of 6 skills)	Minimal-code discipline, always-on	writing from scratch each time	Enforces a YAGNI decision ladder before any code is written. Self-contained, no hooks, no background cost.
taste-skill (1 of 13 skills)	Anti-slop frontend/UI discipline, always-on for frontend work	Anthropic's frontend-design plugin	79.8k stars vs. 33.9k (repo-wide, not plugin-specific). More variants, adjustable dials, and works without /plugin access. See comparison below.
superpowers (13 of 14 skills)	TDD, systematic debugging, planning, explicit-call	get-shit-done-only workflows, ruflo	Anthropic-marketplace-accepted, mature.
code-review-graph	Local code index, MCP server	graphify, sigmap	Tree-sitter plus SQLite, no LLM or embeddings needed. Claims ~65x token reduction on review questions. Narrower scope than graphify (code-only vs. code+docs+PDFs), which is exactly the fit for day-to-day review work.
graphify	Cross-format knowledge graph (code+docs+PDF+SQL), explicit-call only	keeping it always-on	Broader scope than code-review-graph, but that breadth means redundant indexing if run alongside it. Installed dormant: invoke by name only when a task is genuinely cross-format.
markitdown (Microsoft)	PDF/DOCX/PPTX to Markdown	building a custom converter	Official, low-risk, single-purpose. No overlap with anything else here.
context-mode	Sandboxes large tool output into summaries	caveman, Paritok	Broader hook surface (Bash/Read/Grep/WebFetch/Task) than any competitor in this niche. Installed as MCP tools only; hook-based auto-interception is opt-in, see note below.
get-shit-done (community fork)	Milestone/phase planning, 31 explicit /gsd:* commands	the original get-shit-done-cc npm package	The original is marked unsupported/abandoned on npm. This installs conradvc/CC-get-shit-done, an actively maintained fork. Verify it still works before depending on it long-term.
rtk	Bash output compression	Paritok, caveman, token-optimizer	77k+ GitHub stars, 16-tool integration, most battle-tested option in this category. See swap note below.
claude-token-efficient rules	Response verbosity rules in CLAUDE.md	writing custom rules	6k+ stars, single-file, zero install risk, directly additive.

What we left out of the source repos, on purpose

This installs the useful subset of each project, not the full repo. Two examples where that mattered:

ponytail ships 6 skills total: the core skill, ponytail-review, ponytail-audit, ponytail-debt, ponytail-gain, and ponytail-help. This installs the first 4 and skips the last 2. ponytail-gain only prints a static benchmark scoreboard (fixed numbers from the project's own README, not measured from your repo). ponytail-help only prints a reference card. Neither changes any behavior, so neither earns a place in an always-on install.

superpowers ships 14 skills. This installs 13 and drops using-superpowers, a meta-skill whose own instructions say to invoke any skill that's even 1% relevant before responding to anything, including clarifying questions. That's the opposite of the explicit-call model this repo is built around, so it's excluded even though the other 13 skills from the same project are kept.

taste-skill ships 13 skill variants: the core design-taste-frontend skill, a preserved design-taste-frontend-v1, a GPT/Codex-optimized version, image-to-code and redesign workflows, several stylistic variants (minimalist, brutalist, high-end-visual-design), image-generation skills, and more. This installs 1: the core design-taste-frontend variant. Running all 13 as always-on would mean multiple overlapping frontend-design skills all trying to weigh in on the same task, with no clear priority between them. Install a specific variant yourself (npx skills add https://github.com/Leonxlnx/taste-skill --skill "<name>") if you want a different one instead of, not in addition to, the default.

What's deliberately not installed

Tool	Why skipped
ruflo	Multi-agent orchestration framework built for teams running always-on swarms. Wrong shape for individual work, no clean "dormant until needed" mode.
claude-mem	Redundant with Claude Code's own native cross-session memory system. Adds compression/injection token overhead for something already free.
caveman, Paritok, token-reducer (madhan230205)	Each overlaps an installed tool (context-mode or code-review-graph) with a smaller, less proven track record.
sigmap	Genuinely useful verify (anti-hallucination) feature, but its indexing overlaps code-review-graph. Worth adding only if you actually hit a hallucination problem, not preemptively.
Original get-shit-done-cc (npm)	Marked unsupported/abandoned by its own maintainer.
Anthropic frontend-design plugin	Chose taste-skill instead: more variants, adjustable dials, no /plugin requirement. See docs/COMPARISONS.md for the full breakdown.
The other 12 taste-skill variants	Only design-taste-frontend installs by default; see "What we left out" above for why running all of them isn't the goal.

Best combinations

There isn't one right setup. Below are three, and the reasoning for which one this repo defaults to.

Combination	What's in it	Best for	Tradeoff
1. Default (what this repo installs)	ponytail, taste-skill, superpowers, code-review-graph, context-mode (MCP only), rtk, graphify, markitdown, get-shit-done	Solo developers, no /plugin access, want it working today	rtk's automatic hook needs the standard claude CLI process; some hosted environments won't run it (see note below), leaving you on manual invocation until you have that access
2. Token-max	Same as default, but swap rtk for token-optimizer, and wire context-mode's Bash/Read/Grep/WebFetch hooks fully	Anyone with confirmed /plugin access who wants the broadest possible token coverage	token-optimizer's biggest savings numbers are self-reported and unaudited; only a smaller subset is independently metered. More hook surface also means more that can silently misfire
3. Team / verified	Default plus sigmap (for sigmap verify, catching hallucinated file or symbol references)	Teams or anyone burned by an AI assistant inventing a file/function that doesn't exist	Redundant indexing with code-review-graph for the base "answer questions about my code" job; you're paying disk and build time twice for the 80% both tools already do

Why this repo defaults to Combination 1: it's the only one of the three that works without any precondition. It doesn't assume /plugin access (which isn't available in every Claude Code environment, including some hosted/SDK-based sessions), doesn't assume you've already hit a hallucination problem worth sigmap's overhead, and every tool in it has an independently proven track record rather than a projected one. It's also the lightest: no duplicate indexing (one code-index tool, not two), no fully wired hook surface (context-mode's broader hooks stay off by default), and nothing that requires a background service. Combinations 2 and 3 are real upgrades for specific situations, not corrections to the default.

Swap options

Want token-optimizer instead of rtk? Run this instead of the rtk step (needs /plugin access):

/plugin marketplace add alexgreensh/token-optimizer
/plugin install token-optimizer@alexgreensh-token-optimizer

Then remove rtk's hook entry from ~/.claude/settings.json so they don't both intercept Bash at once, or run install.sh --skip-rtk from the start so rtk is never installed in the first place.

Want sigmap alongside code-review-graph? Install it separately; this repo doesn't include it. Only worth adding if you specifically want sigmap verify, checking answers against your actual code to catch hallucinated files or symbols. Running both for the base indexing job is redundant, not complementary.

Important: the rtk hook may not auto-fire everywhere

rtk's automatic bash compression depends on Claude Code's PreToolUse hook mechanism, which requires the standard claude CLI process reading ~/.claude/settings.json. Some hosted/SDK-based Claude Code environments (certain desktop apps, cloud sessions) don't run that process and won't honor the hook even though rtk init -g reports success. If a verbose command comes back uncompressed after installing, this is why. The fix is either running from an actual claude CLI terminal, or invoking rtk <command> manually for large-output commands.

What "install" actually touches

• ~/.claude/skills/: ponytail (4 of 6), taste-skill (1 of 13), superpowers (13 of 14)
• ~/.claude/commands/gsd/: 31 command files
• ~/.claude/get-shit-done/: supporting workflows/templates/bin
• ~/.claude/CLAUDE.md: appends rule sections (ponytail, taste-skill, token-efficiency), does not overwrite existing content
• ~/.local/bin/: rtk binary
• pipx-isolated venvs: code-review-graph, graphify, markitdown (no shared dependencies with your other Python projects, deliberately, see the note below)
• npm global: context-mode
• .mcp.json in the directory you ran the installer from

Why pipx, not plain pip

A plain pip install puts these tools' dependencies into your global Python environment, which can silently upgrade packages other projects depend on (this repo exists partly because that happened during initial testing: starlette, pydantic, and websockets got bumped and broke unrelated projects in the same environment). pipx gives each CLI tool its own isolated venv instead.

More docs

• docs/INSTALL.md: platform-specific install steps
• docs/UNINSTALL.md: removing everything cleanly
• docs/TROUBLESHOOTING.md: common problems and fixes
• docs/TOOLS.md: every tool, in depth
• docs/COMPARISONS.md: the evidence behind each "chosen over" call
• SECURITY.md: what this script does, what it doesn't, how to inspect it first
• CHANGELOG.md: what changed and when

License

MIT, see LICENSE. Individual tools installed by this script retain their own upstream licenses; nothing here is vendored or relicensed.