Other routes exist, and the documentation lists all three:

The installer can install piecemeal, show what it intends to do, and skip confirmations:

MCP servers install separately and selectively:

Something useful that neighbouring projects often lack: a health check and backups.

The Price Is Context

This is the most important section, and the figures come from the documentation itself.

The "what's installed" table in the installation guide lists: the ~/.claude/ directory at around 50 MB, the entry point ~/.claude/CLAUDE.md at around 2 KB, MCP settings at around 5 KB - and the behavioral instruction files ~/.claude/*.md at around 200 KB.

Two hundred kilobytes of Markdown is, at the usual ratio for English text, tens of thousands of tokens. I recounted what feeds those files from the repository: the base rules and principles files come to about 49 KB, and the seven mode files add about 27 KB. The largest single one is RULES.md at 16 KB.

In the project's discussions there is a question raised back in October 2025: SuperClaude eats about twenty thousand tokens, is there a way to disable it and use it only when needed. On the page as viewed it remained unanswered. Beside it sits an equally unanswered question: is SuperClaude always active, on every prompt.

The project gives no direct answer, but the interface offers a partial way out. Installing by component lets you skip what you do not need, and one of the seven modes is called exactly that - token efficiency - with a claimed 30 to 50 percent context saving in the README. A saving tool inside a tool that itself costs context.

Behavioral Modes

Seven sets, each for its own kind of work.

The largest by volume is Business Panel at nearly 12 KB of instructions. The most compact is Deep Research at about 1.6 KB.

MCP Is Not Part of the Framework

Here the documentation makes an important qualification, and it is worth reading before installing. The section is literally titled "reality check".

What MCP servers are: external Node.js processes that provide additional tools. What they are not: built-in SuperClaude functionality. That is the wording verbatim.

The eight servers are: Tavily for web search, Context7 for library documentation, Sequential-Thinking for multi-step reasoning, Serena for memory across sessions, Playwright for the browser, Magic for generating interfaces, Morphllm-Fast-Apply for code edits, and Chrome DevTools for performance analysis.

Two of them are paid, and the documentation says so in the troubleshooting section: Magic and Morphllm errors are expected without API keys, these are paid services. Tavily has a free tier. The servers themselves need Node.js 16 or later.

The Claimed 2-3x and 30-50 Percent

The README states: with MCP servers, execution is two to three times faster and uses 30 to 50 percent fewer tokens. It is repeated below as a comparison line: without MCPs, fully functional and standard performance; with MCPs, faster and cheaper.

There is no method behind those numbers at all. No model, no workload description, no run count, no baseline. All there is is a note that Serena is responsible for the speed and Sequential for the token saving.

The research-depth figures, by contrast, are stated properly, because they are settings rather than measurements: up to five iterative searches, 20 to 40 sources at the deep level, and over 40 at the exhaustive one.

The Documentation Disagrees With Itself

The discrepancies are not small, and knowing about them saves you hunting for things that do not exist.

The project's front page claims 30 commands, 20 agents, 7 modes and 8 MCP servers. The documentation index in the docs/ directory of the same tree says 6 MCP servers.

The plugin's own README - a two-page file - promises 7 skills in a subheading, while the table directly under that heading lists six: confidence check, deep research, brainstorm, troubleshoot, project management and token efficiency. There is no seventh in the directory - I checked, there are exactly six subdirectories. The same file's MCP section lists two servers instead of eight.

The Python requirement has drifted: the installation guide says 3.8 and later, while the package metadata requires 3.10 and later. The second is correct - otherwise version 4.3.0 simply will not install.

Release 4.3.0 separately records that a documentation overhaul was done, aligning versions and accurate feature counts. Some pages, by the look of it, never got there.

The Plugin Exists, the Marketplace Does Not

Care is needed here, because instructions circulating online do not work.

Installing through Claude Code's plugin marketplace is shown in the README, but with a note: this feature is not yet available. Status: in development, no ETA set, planned for version five. Accordingly, the whole TypeScript plugin system from older documentation does not exist either.

The plugins/superclaude directory does exist in the repository and does work - but locally, by launching Claude Code with a directory pointer:

And right beside it a warning that is easy to miss: the pip and npm versions are not compatible with the plugin version, and installing them at the same time is not allowed.

What to Know in Advance

There has been no release for five months. The last one is 4.3.0 from 22 March 2026, and that is also what sits on PyPI. Commits, however, keep coming: the latest is from 21 August 2026. The repository is alive; the release train is frozen.

Copies of the files have to stay in sync. Commands, agents and hooks live in two places - src/superclaude/ for the package and plugins/superclaude/ for the plugin. The requirement is stated in their READMEs in plain words: both locations must stay in sync. If you edit your own copy, edit both.

Release 4.3.0 closed holes rather than adding features. From its description: a shell=True command execution vector was removed, SHA-256 integrity verification was added, and three placeholders in the code - the confidence checker, intelligent execution and the reflexion pattern - were replaced with real implementations. If you are on something older, upgrade for that reason rather than for new commands.

Hooks attach to three events. Session start, stop, and file write: on stop it checks for uncommitted changes and unfinished tasks, on write it verifies the edit.

The framework has its own test suite, but the 136 tests claimed in the release are test cases, not files. There are ten files in the two main directories.

First-Encounter Mistakes

Expecting software. There is none. There is a set of Markdown files that changes Claude Code's behaviour, and all execution stays inside Claude Code.

Installing the package and thinking you are done. Installing the package and superclaude install are two separate steps.

Following instructions that use the plugin marketplace. It does not work; it is planned for version five with no date.

Installing the pip version alongside the plugin version. The project warns outright that they are incompatible.

Believing the 3.8 in the installation guide. The package requires Python 3.10 or later.

Looking for the seventh skill. There is none, though the plugin README promises seven.

Treating MCP servers as part of the framework. They are external Node.js processes, and two of the eight require paid keys.

Taking 2-3x and 30-50 percent as a measurement. There is no method behind those numbers in the documentation.

Installing everything without thinking about context. Two hundred kilobytes of instructions in ~/.claude/ is a figure from the project's own documentation, and users are not complaining about twenty thousand tokens for nothing.

To put it in one sentence: SuperClaude buys you structure and a vocabulary of commands, and you pay for it with your context window - so the decision comes down to whether that structure is worth its standing price in tokens.

Sources

This article was checked against the SuperClaude-Org/SuperClaude_Framework repository (branch master) on 25 August 2026: the README, pyproject.toml, plugins/superclaude/README.md, docs/README.md, docs/getting-started/quick-start.md, docs/getting-started/installation.md, docs/user-guide/mcp-servers.md, and the plugins/superclaude/ and src/superclaude/ trees. The command, agent and mode counts were recounted from the tree and matched the claimed figures; the skill count did not. Instruction file sizes come from the repository tree, and the 200 KB figure from the project's own documentation. Star counts were taken the same day. The last release shipped in March 2026, so documentation discrepancies have had time to accumulate.

• GitHub: SuperClaude-Org/SuperClaude_Framework
• Quick start
• Installation
• Command reference
• MCP servers
• The plugin README
• The documentation index
• Release 4.3.0
• The package on PyPI
• Project discussions