Everyone Runs Claude Code on Default Settings. Here’s How to Stop.
Five extensions, some MCP servers, some Skills, that change what Claude Code can actually do for you.
Claude Code is good enough out of the box that most people never go looking for more. You install it, you point it at a repo, it writes code, you ship. That’s the whole loop for 90% of users right now.
But “good enough” and “max potential” are different products. The gap between them isn’t a smarter model, it’s the five or six extensions almost nobody installs because nobody told them these existed, what category they fall into, or why it matters which category they’re in.
That last part actually matters, so let’s get it out of the way first.
MCP vs. Skills: the two upgrade paths
Claude Code extends in two fundamentally different ways, and conflating them is why half the setup guides online are confusing.
MCP servers give Claude new capabilities, actual tool calls to actual running processes. An MCP server for image generation means Claude can call a function that returns a real PNG. An MCP server for token compression means Claude can call a function that shrinks a 50,000-token log file before it ever touches the model. You install these with claude mcp add, they run as a background process or remote service, and Claude invokes them like any other tool.
Skills give Claude new behavior, instructions, not capabilities. A skill is a markdown file (often called SKILL.md or CLAUDE.md) that gets loaded into context and tells Claude how to think about a class of problem: when to use it, what rules to follow, what the desired output looks like. No server, no API calls. Just a sharper set of defaults.
Some of the five tools below are MCP. Some are Skills. One does both depending on which install path you pick. Knowing the difference tells you what you’re actually signing up for — a new tool, or a new set of habits.
1. Higgsfield, give Claude eyes and hands for visual work
Higgsfield is the odd one out on this list because it’s not really about coding, it’s about everything coding adjacent to a launch. Marketing assets, hero images for the README, product mockups, UGC-style video for a landing page. Normally that means tab-switching to a separate app, generating, downloading, re-uploading. Higgsfield’s MCP server collapses that into the same terminal session.
It’s an MCP server, hosted by Higgsfield directly, and the setup is refreshingly free of the API-key dance most generation tools force on you:
claude mcp add --transport http --scope user higgsfield https://mcp.higgsfield.ai/mcpFirst tool call triggers a browser OAuth flow against your Higgsfield account, sign in once, and every Claude Code session on that machine shares the connection. No keys to rotate, no secrets in your .env.
Once it’s wired in, Claude has access to 30+ underlying models, Soul, Cinema Studio, Flux, Seedream, Kling, Veo, GPT Image 2, and picks the right one for the job, or you specify it yourself. Train a consistent character once and reuse it across a whole campaign. Drop in a reference video and have Claude reverse-engineer its pacing and shot structure into a new prompt. There’s also a dedicated CLI for people who live entirely in the terminal, and a layer of community-built prompt-engineering skills on top that turn a vague “make me a hero image” into the five-field brief, model, composition, subject, lighting, aesthetic, these models actually respond to.
The MCP gives Claude the ability to generate. The prompt-engineering skills are what make the generations not look like everyone else’s.
2. Ponytail: the senior dev who refuses to write 200 lines
Anyone who’s used Claude Code on a real codebase has hit this: you ask for a date picker, and your agent installs a dependency, writes a wrapper component, adds a stylesheet, and starts a paragraph about timezone edge cases. Ponytail exists to stop that.
It’s built around a “ladder” the agent has to climb before writing code: does this need to exist at all? Is it already in the codebase? Can the standard library do it? Only after exhausting those does it write something new — and it’s instructed to write the minimum, not the impressive version.
Ponytail ships primarily as a Skill, paired with two lightweight lifecycle hooks (SessionStart and UserPromptSubmit) that inject the ruleset automatically and track mode changes. There’s also a standalone ponytail-mcp server for hosts that don’t support hooks or skills natively. Inside Claude Code:
/plugin marketplace add DietrichGebert/ponytail
/plugin install ponytailReview and trust the two hooks when prompted, and you’re running with it from the next session on.
The benchmark behind this is worth knowing because it’s unusually honest. The original number, “80–94% less code”, was a single-shot comparison that a critic pointed out was partly an artifact of the baseline being chatty. The project’s response was to rerun the test properly: a real headless Claude Code session, editing a real open-source FastAPI + React repo, twelve feature tickets, with and without the skill. The honest number that came out the other side: ~54% less code on average, ~20% cheaper, ~27% faster, with every safety check still intact. It’s a smaller number than the marketing version, and it’s more believable for it.
3. Headroom: make your token budget last four times longer
If you’ve ever watched Claude Code chew through a context window on a single debugging session because a log file or grep dump was 40,000 tokens of mostly noise, Headroom is the fix. It sits between your agent and the model and compresses everything the agent reads, tool outputs, logs, RAG chunks, files, even prior conversation turns, before billing happens. The published numbers are 60–95% fewer tokens with no drop in answer quality, and compression is reversible: the model can pull the original back via headroom_retrieve if it actually needs the verbatim text later.
Headroom is unusual in that it works across every category we’ve talked about: a Python/TypeScript library, a transparent proxy, a CLI wrapper, and an MCP server exposing three tools, headroom_compress, headroom_retrieve, headroom_stats. Simplest path:
pip install “headroom-ai[all]”
headroom mcp install
claudeFor automatic compression of everything (not just on-demand calls), run it as a proxy instead and point Claude Code at it:
headroom proxy --port 8787
ANTHROPIC_BASE_URL=http://127.0.0.1:8787 claudeOne thing worth knowing before you flip this on blindly: a user reported that running Headroom purely in MCP mode during a long deep-research session ate 25% of their usage limit, because MCP tool results themselves sit in context for the rest of the session, compression doesn’t help if the compressed-but-still-large result then gets read repeatedly. The proxy mode, which compresses transparently at the request layer, is the safer default for raw token savings; MCP mode is better thought of as a tool Claude reaches for selectively, not something to leave running unsupervised on huge sessions.
4. “Andrej Karpathy’s skills” and the part most posts get wrong
This one needs a correction before it needs a setup guide, because the name oversells the provenance. Andrej Karpathy did not write a skills file. In January 2026, he posted a list of observations about how LLM coding agents fail, they run with wrong assumptions instead of flagging them, they bloat 100 lines of logic into 1,000, they touch code they don’t understand as collateral damage. Developer Forrest Chang read those observations and turned them into a single structured file, hosted at forrestchang/andrej-karpathy-skills, which is where the actual engineering happened.
That file, just CLAUDE.md, no dependencies, no model, has somehow become one of the most-starred repos in GitHub’s recent history. It’s a pure Skill: four behavioral principles (Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution) loaded straight into context, no server involved. Install it as a Claude Code plugin so it applies across every project:
/plugin install andrej-karpathy-skills@karpathy-skillsOr drop it directly into one project’s root:
curl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.mdKarpathy diagnosed the disease. Forrest Chang wrote the prescription. The internet just keeps crediting the wrong doctor.
The genuinely useful idea buried in it, “don’t tell the agent what to do, give it success criteria and let it loop” is the same principle Ponytail and Headroom’s headroom learn both lean on in different ways. Verification loops, not instructions, are what 2026’s better agent tooling has converged on.
5. Markitdown: feed Claude literally anything
The least flashy tool here is the one you’ll use the most once it’s installed. MarkItDown is Microsoft’s open-source converter for turning PDFs, Word docs, PowerPoint decks, Excel sheets, images (with OCR and EXIF extraction), audio (with speech transcription), HTML, CSV/JSON/XML, EPUB, and even YouTube transcripts into clean Markdown. The reasoning behind picking Markdown as the output format is simple: it’s what LLMs were trained on the most of, it’s dense with structure, and it’s cheap in tokens compared to raw HTML or binary-derived text.
Microsoft ships an official MCP server for it, markitdown-mcp, which is the cleaner of the two integration paths if you want it always available:
claude mcp add markitdown -- uvx markitdown-mcpZero config, no API keys, it just exposes one tool, convert_to_markdown, and Claude calls it whenever a file or URL shows up that needs converting. If you’d rather not keep a server process running, the community has also wrapped the same library as a Skill that shells out to the markitdown CLI directly, better for occasional or batch conversion jobs where spinning up a server feels like overkill.
One limit worth knowing before you rely on it for a stack of scanned contracts: MarkItDown extracts structure, headings, tables, lists, links — not visual fidelity, and there’s no built-in OCR on the PDF path specifically. A scanned PDF with no text layer comes back empty. OCR only kicks in on the image-conversion path. For genuinely scanned documents you’ll want Docling or an Azure Document Intelligence endpoint, both of which MarkItDown can hand off to if you configure it.
Stacking them
None of these five are doing the same job, which is exactly why they’re worth running together instead of picking one. Markitdown is your ingestion layer, anything that lands on your desk becomes something Claude can actually read. Headroom is your cost layer — it decides how much of what gets ingested actually has to hit the model. Ponytail and the Karpathy CLAUDE.md are your output layer, they’re both, in different words, telling Claude to write less and verify more. And Higgsfield is the one that has nothing to do with code at all, which is the point: it’s there for the moment your project needs a visual and you don’t want to leave the terminal to get one.
The honest takeaway isn’t “install all five today.” It’s that Claude Code was designed to be extended this way from the start, MCP for new capabilities, Skills for new judgment, and the people getting noticeably more out of it right now aren’t using a different model. They’re just the ones who bothered to look.
Sources: