SuperClaude Framework
A configuration layer installed on top of claude-code that ships 30 slash commands, 20
specialized agents, 7 behavioral modes, and wiring for 8 MCP servers (Tavily for search,
Context7 for documentation lookup, Playwright for browser automation, and others). MIT, 23.7k stars,
current stable v4.3.0. It installs as a Python package — pipx install superclaude then
superclaude install — which writes the configuration into the harness rather than running as a process
of its own. The project describes itself as a “meta-programming configuration framework” that works by
behavioral instruction injection: the whole thing is text placed where the agent will read it.
Tier: T3 — the project’s own README. Star count aside, nothing here is independently measured, and the framework makes the kind of claim (“transforms Claude Code into a structured development platform”) that only an evaluation could settle.
Where it sits
This is the same species as superpowers — a packaged development methodology a coding harness reads and follows — and the two are worth reading against each other, because they package it differently.
| superpowers | SuperClaude | |
|---|---|---|
| Unit | composable skills the agent reaches for on its own (agentskills-spec progressive disclosure) | slash commands the operator types, plus modes and agents |
| Harness reach | cross-harness (Claude Code, Cursor, Codex, Copilot CLI, Gemini CLI…) | Claude Code |
| Install | plugin marketplaces | pipx, writes config in |
| Tooling | none bundled | 8 MCP servers wired in |
The interesting difference is who decides. Superpowers bets on the agent recognizing that a situation
calls for brainstorming or systematic-debugging and loading it unprompted. SuperClaude’s primary
surface is a command the human invokes — /brainstorm, /implement, /research, /test, /pm — with
the 7 behavioral modes as the standing-state half. The phase vocabulary is nearly identical (both have a
brainstorm step, both have research, both have test); the control point is not. That is the
operator-triggered vs agent-triggered axis showing up inside two projects that otherwise want the same
thing.
Its 20 agents are role personas — PM Agent, Deep Research Agent, Security Engineer, Frontend Architect — which puts it beside auto-company‘s 14 personas rather than beside orchestration systems that split work by task (orchestrate-100-agents-claude-code). Personas as a division of labour keep recurring in this corpus and remain unmeasured against a plain agent with a good prompt.
v5 is in development as a TypeScript plugin system, which would move it off the pipx-writes-config distribution and onto the same plugin rails superpowers already uses.
What is not here
No evaluation. 23.7k stars measure that people install it, which is the same limit the corpus already recorded for cc-thinking-skills — and that was the one case where somebody actually ran the harness and found zero proven accuracy gains across 39 similarly-motivated skills. A framework whose entire mechanism is injected instruction text is exactly the thing that ablation applies to. Until someone runs it, SuperClaude’s 30 commands sit on the untested side of the line context-files-ablation-paper drew.
Related
superpowers · claude-code · agentskills-spec · agent-orchestration · spec-driven-development · cc-thinking-skills · auto-company · model-context-protocol