.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
Search for a "super" version of Claude Code and you land on SuperClaude, a framework with over 23,000 GitHub stars that adds 30 slash commands, 20 specialist agents and 7 behavioral modes to Anthropic's CLI. Worth getting one thing straight before installing anything: SuperClaude works by instruction injection. The installer writes a set of behavioral instruction files into your Claude Code configuration, and from the next session on the same binary reads those files and acts differently. There's no fork involved and no proxy sitting between you and the API.
I spent a few evenings with it on an Ubuntu 24.04 machine running the 2.1 line of the CLI. This article covers the install, what a session feels like afterwards, what the injected instructions cost you in context and my honest read on who comes out ahead using it. The SuperClaude repository is active, with v4.3.0 released in March 2026 and a v5 in development.
What SuperClaude adds to Claude Code
Four things, concretely. Thirty slash commands under an /sc: prefix, running from /sc:brainstorm for requirements discovery through implementation and testing to /sc:research for deep web research. Twenty agents framed as personas, a PM, a Security Engineer, a Frontend Architect and colleagues, each one a bundle of instructions about how that specialist approaches work. Seven behavioral modes that change how the agent operates, including Brainstorming, Deep Research, Orchestration and a Token-Efficiency mode (more on that one later, it's a tell). And eight optional MCP servers: Tavily for web search, Context7 for library docs, Serena, Playwright, Magic, Chrome DevTools, Morphllm-Fast-Apply and Sequential, which is the same sequential thinking server I've covered on its own.
Every one of those is text. The commands are prompts, the personas are prompts, the modes are prompts that tell Claude when to act differently. That's worth repeating because it sets expectations correctly: SuperClaude can't do anything Claude Code couldn't do with the same instructions pasted in by hand. What it sells is the design work of writing, organizing and shipping those instructions as a coherent set.
Install SuperClaude with pipx
The documented path is the Python package via pipx, which keeps it out of your system Python:
pipx install superclaude
superclaude installThe second command writes the framework's components into your Claude Code setup. Two checks worth running straight after: superclaude install --list shows what landed and superclaude doctor reports on the installation's health. Then restart Claude Code, because instruction files load at session start. Typing /sc in a fresh session should now offer the framework's commands in the completion menu.
Optional MCP servers with superclaude mcp
The MCP servers are opt-in and there's a picker:
superclaude mcp --list
superclaude mcp --servers tavily --servers context7Plain superclaude mcp walks you through the full set. These are ordinary MCP servers using the same plumbing as any server you'd wire up yourself through Claude Code's MCP configuration, and some bring their own requirements, Tavily's web search wants its own API key for instance. My advice is to skip the bulk install and add the two or three you'll genuinely use. Every connected server contributes tool definitions to your context, and eight of them at once is a lot of furniture.
What a session looks like with SuperClaude
Run /sc:brainstorm with a vague product idea and Claude starts interviewing you about requirements instead of writing code, which is the Brainstorming mode doing its job. Ask for an implementation and the framework's routing instructions nudge the model to lean on the relevant persona, so a payment feature gets security-flavored pushback and a UI task gets frontend opinions. The modes switch on cues in your messages and the current task, so the same request can play out differently depending on how you phrase it.
The overall effect is pair programming with someone who insists on process. For greenfield work that discipline can be useful, and the command vocabulary gives structure to people who'd otherwise type "build me the thing" and hope. For a two-line fix it's ceremony, and you'll find yourself working around your own tooling. Both reactions are common and both are fair.
Context cost of injected instructions
Here's the part the README doesn't dwell on. Everything SuperClaude injects occupies context-window tokens in every session before you've typed a word. Commands, personas, mode definitions, all of it rides along, and on top of that each optional MCP server adds its tool schemas. The framework shipping a dedicated Token-Efficiency mode, advertised as saving 30 to 50 percent of context, tells you the authors treat context spend as a real problem in their own system. I won't quote an overhead number because it depends entirely on which components you install; run /context in a session before and after installing and read the difference off your own setup.
Why care? Context spent on framework instructions is context unavailable for your code, and if you're on a subscription it feeds the same session budget as everything else, which is the mechanism behind the usage limits I've written about separately. On an API key it's simply tokens billed on every request. A framework has to earn its rent per session, and that's the honest lens for the next section.
SuperClaude vs plain Claude Code
Almost everything SuperClaude packages maps onto a stock feature. Project conventions and process belong in a CLAUDE.md, which loads at session start and can carry exactly the instructions you want and nothing else. Repeatable procedures are skills, loaded on demand instead of permanently. Personas are a description field away from being custom subagents, with the advantage that a subagent runs in an isolated context and can be restricted to specific tools, which a persona prompt can't. And the MCP servers install fine without any framework around them.
So the comparison isn't capability, it's who does the design work. SuperClaude gives you a large, internally consistent instruction set that somebody else built, tested together and maintains. Building the equivalent yourself takes weeks of iteration, and most people's homegrown setups are smaller and less polished. The price is that someone else's opinions now load into every session, and when the agent does something odd, the answer to "why did it do that" might live in one of dozens of instruction files you didn't write. I've debugged enough surprising agent behavior to weigh that cost heavily. A twenty-line CLAUDE.md I wrote myself fails in ways I can find in one read.
SuperClaude v4 and the planned v5
One caution for anyone evaluating it in 2026: the current release line is v4 (v4.3.0 as I write this), and the README describes a v5 built on a TypeScript plugin system, tracked in issue #419, that would reduce installation to a single /plugin marketplace add command. The project's own status on that is in development with no ETA. Judge the framework on what v4 does today, and treat any writeup describing v5 features as describing plans.
Uninstall SuperClaude
The README documents no uninstall command as I write this, which shapes how I'd install it in the first place: put ~/.claude under version control (or snapshot it) before running superclaude install, so removal becomes a diff you can revert instead of an archaeology project. If you skipped that step, superclaude install --list enumerates the installed components so you know what to hunt for, and pipx uninstall superclaude removes the installer tool itself once the injected files are gone. MCP servers it added come out with claude mcp remove <name> like any other.
Who should use SuperClaude
My take after living with it: run vanilla Claude Code first, for weeks, with a CLAUDE.md you grow by hand. Most developers discover the stock tool plus their own small instruction set covers them, and they keep full visibility into what the agent has been told. Reach for SuperClaude when you want opinionated structure without building it yourself: you like the idea of named workflows and specialist personas, you'd rather adopt a system than design one and you're willing to pay the context rent for it. That's a real audience and the star count says it's a big one. If you're still learning the base tool, start with the fundamentals in our guide on how to use Claude Code, because a framework layered on a tool you don't yet know doubles the mystery when something misbehaves.
