What Superpowers Actually Does, at 291,000 GitHub Stars
Software / explainer
What Superpowers Actually Does, at 291,000 GitHub Stars
The plugin makes a coding agent stop and ask what you're building before it writes anything, and its latest release adds support for Meta's Muse.

What is Superpowers?
Superpowers is a free plugin that changes how a coding agent like Claude Code behaves, rather than a coding agent itself. Its GitHub repository describes it as "a complete software development methodology for your coding agents, built on top of a set of composable skills." Created on Oct. 9, 2025 by Jesse Vincent, founder and chief executive of the applied research lab Prime Radiant, it is licensed under MIT and, as of Sept. 25, carries 291,231 stars, 26,055 forks and 401 open issues on GitHub, having gained 611 stars on Sept. 25 alone according to GitHub's daily trending list.
The easiest way to install it inside Claude Code is a single command pulled from Anthropic's own marketplace: /plugin install superpowers@claude-plugins-official. The README also documents installation on more than a dozen other agent platforms, from Cursor and GitHub Copilot CLI to Devin CLI and Google's Gemini CLI.
How does the workflow actually run?
Superpowers replaces a coding agent's usual jump straight to writing code with a fixed sequence of skills. Brainstorming comes first: the agent asks what you're actually trying to build, reflects your answer back for correction, and does not treat "that scope is okay" as permission to start building, according to the changelog for version 6.3.0, published Aug. 12. Writing-plans follows, breaking the work into tasks small enough for one review cycle each. Execution then runs one of two ways: subagent-driven development, which dispatches a separate implementer and a single reviewer for each task, or, since version 6.4.1, a cheaper "Native execution" mode in which the same session does the work itself under the same ledger and stopping rules.
Jesse Vincent's own numbers for the older, more expensive mode are vendor-supplied and not independently verified: the version 6.0.0 release notes, published June 16, say that in Prime Radiant's internal evaluations, "Claude Code and Codex produce similar high-quality results roughly twice as fast and while spending almost 50% fewer tokens" than earlier versions of the same tool. No outside benchmark of that comparison has been published.
Why did it just add Meta's Muse?
Version 6.4.1, published Sept. 19, added support for three more coding agents: OpenCode 2.0.4, Qwen Code, and Muse, the terminal-based coding agent Meta launched on Aug. 5 and that this site covered this week after a tester exported 6.8 gigabytes of its own runtime files to Google Drive. Each new "harness," in the project's terminology, gets its own plugin manifest and its own session-start hook so the tool's skills load the same way regardless of which agent is running them.
| Recent Superpowers releases | Date | What it added |
|---|---|---|
| v6.1.0 | June 30, 2026 | Codex marketplace install; Gemini CLI support removed |
| v6.2.0 | July 24, 2026 | Plan-scoped workspaces; resumable review-fix loop |
| v6.3.0 | Aug. 12, 2026 | Devin CLI and Hermes Agent support |
| v6.4.1 | Sept. 19, 2026 | OpenCode 2.0, Muse and Qwen Code support; new diagnosing-superpowers skill |
The diagnosing-superpowers skill added in the same release is meant to answer a narrower question: when something goes wrong in a session, it reads the transcript on disk and reports what happened with a file-and-line citation for every finding, rather than asking the developer to reconstruct it by hand.
Who is verifying any of this?
The clearest outside account of Superpowers comes from Simon Willison, an independent software developer who writes about coding agents and who profiled the plugin on Oct. 10, 2025, the day after Vincent released it. Willison called Vincent "one of the most creative users of coding agents" he knew and described the piece documenting Superpowers as "a really significant piece in its own right," pointing specifically to Vincent's use of true red/green test-driven development and what Willison called self-updating memory notes.
A second outside account, an April 23, 2026 episode of Heavybit's Open Source Ready podcast, put the plugin at roughly 115,000 GitHub stars and over 200,000 installs at that point, five months after launch, already the most-installed third-party Claude Code plugin. The repository has since grown to 291,231 stars, an increase of more than 150% in five months, according to GitHub's own API.
- Apr. 23, 2026115K stars
- Sept. 25, 2026291K stars
Source: Heavybit Open Source Ready podcast (Apr. 23, 2026) and GitHub API (Sept. 25, 2026)
Neither outside account tests Vincent's specific cost or speed claims; both describe the plugin's design and its reach, not its accuracy. Prime Radiant's own contributor guidelines, added in the version 5.1.0 release notes on May 4, disclose a separate number worth weighing against the star count: an audit of the project's last 100 closed pull requests found a 94% rejection rate, which Vincent attributed to "AI-generated slop," agents that skipped the PR template, opened duplicates, or fabricated the problem they claimed to be fixing. Superpowers is built to change how an agent works inside one codebase. Getting a contribution into the codebase that builds it is apparently a different problem.
Superpowers is one entry in the same wave of GitHub-trending agent tooling that produced CLI-Anything, a project this site found making broader claims than its own technical documentation supported. Vincent's changelog is unusually specific about what changed and when, which is not the same as those changes being measured by anyone outside Prime Radiant.
Sources
More in Software
- 01Polyend Publishes Tracker Firmware Source Under a Noncommercial Licence, With Two Binary Libraries Left OutThe tracker-ce repository covers the original Tracker, Tracker+ and Tracker Mini, but the synth and DSP libraries ship precompiled under a separate proprietary licence.
- 02Claude-Mem's Installer Uploaded 30 Days of Prompts to cmem.ai Before Asking; v15.8.3 Adds a Consent PanelThe fix shipped on Nov. 2, but the release notes say prompts already uploaded by earlier versions are not removed, and the GitHub issue was reopened the same day.
- 03Vercel Labs' Agent-Desktop 1.0 Ships With a macOS Bug That Types Into the Frontmost AppThe Rust tool lets an AI agent drive desktop apps through the accessibility tree, and 48 open issues sit against it three days after the 1.0.2 release.
- 04T3 Code Reaches 38.1k Stars in Alpha, and Its Remote Access Docs Say to Treat Pairing Links as PasswordsThe MIT-licensed control surface for Claude Code, Codex and other agents has nearly doubled its star count since August 27. Its own documentation lists what stays risky.