Superpowers Reaches 92K Stars: Now Works with Cursor, Codex, OpenCode, and Gemini CLI
By Devy · yoDEV.dev
A few weeks ago we covered Superpowers — the Claude Code plugin that stops the agent before it jumps straight to code. At that time it had around 27,000 stars on GitHub and ran exclusively within Anthropic’s official marketplace.
A lot has changed since then.
Superpowers surpassed 92,000 stars on GitHub, positioning itself as one of the fastest-growing developer tools in 2026. But the biggest change isn’t the number: the framework is no longer exclusive to Claude Code. Jesse Vincent expanded support to Cursor, Codex, OpenCode, and Gemini CLI — turning Superpowers from a Claude plugin into a universal agentic development methodology.
The main shift: cross-platform support
This is the most significant news. Superpowers now installs differently depending on which agent you use:
Claude Code (no changes, still the easiest path):
/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace
Cursor:
/add-plugin superpowers
Codex:
git clone https://github.com/obra/superpowers.git ~/.codex/skills/superpowers
Then follow the instructions in .codex/INSTALL.md within the repo.
OpenCode:
git clone https://github.com/obra/superpowers.git ~/.agents/skills/superpowers
Then follow .opencode/INSTALL.md.
Gemini CLI:
Skills are activated through Gemini’s activate_skill tool. The framework loads metadata at session startup and activates content on demand. There’s a reference file (gemini-tools.md) that translates Claude Code tool names to their Gemini equivalents — for example, Read → read_file, Write → write_file, Edit → replace. Note: Gemini CLI doesn’t support subagents, so implementation skills fall back to executing-plans mode.
An important detail: skill files are now completely portable across platforms. They don’t contain environment variables specific to any agent — the framework adapts to the environment where it’s running. This means if two developers on your team use different tools (one on Cursor, another on Claude Code), they can share exactly the same Superpowers configuration in the repository.
The new skills worth knowing about
The original Superpowers included brainstorming, TDD, systematic debugging, and basic subagent dispatch. The framework has grown considerably. The most useful additions:
using-git-worktrees
Now a required skill before any implementation workflow. Creates an isolated workspace on a new branch, runs project setup, and verifies a clean test baseline before touching a single file. Prevents the most common failure mode: agents working directly on main and leaving a mess.
dispatching-parallel-agents
Runs multiple agents simultaneously in isolated workspaces. Each agent handles a separate task — one writes tests, another implements, another reviews — with context isolation between them to prevent context window contamination. Each agent receives only the context it needs.
efficient-multi-agent-research
Skill contributed by the community (PR #724) for research tasks that benefit from parallelism: multiple agents gather information from different sources simultaneously and then synthesize the findings.
requesting-code-review
Launches a dedicated reviewer agent to review the work before it reaches the main branch. Different from the review gate in the general workflow — you can invoke this at any point during the session.
validate-plans
Validates the implementation plan before execution starts. Detects ambiguities and missing context before subagents start working on tasks that might need adjustment.
The security tools
As Superpowers started being used on larger codebases, the need to control what the agent can and can’t touch became evident. There are three new tools for this:
/careful — Adds explicit warnings before executing any destructive command.
/freeze [directory] — Restricts all edits to a specific directory. Any modifications outside that boundary are blocked. Accident prevention when you need agents working on a specific module without touching adjacent code.
/guard — Combines /careful + /freeze. Use it when debugging production systems or working with sensitive directories.
These aren’t optional suggestions. They’re applied at the session level through Claude Code hooks.
The updated complete workflow
With the new skills, a complete Superpowers session for a non-trivial feature looks like this:
/brainstorming— Refine the idea through structured dialogue before any code/plan— Break the approved design into 2–5 minute tasks with exact file paths/using-git-worktrees— Isolate the workspace on a new branch/dispatching-parallel-agentsor/executing-plans— Run the implementation (parallel or sequential)/requesting-code-review— Dedicated review before merging/retro— Analysis of commits and shipping velocity metrics
The central principle of the framework hasn’t changed: mandatory workflows, not suggestions. Skills activate automatically based on context — you don’t invoke most of them manually. When the agent detects you’re building something, brainstorming activates. When a plan is approved, git worktrees activates. The agent follows the methodology, not ad-hoc prompts.
Why 92K stars makes sense
In 2026, the bottleneck in AI-assisted development isn’t code generation — it’s reliability. Agents are fast, but they deviate, skip tests, make assumptions. Superpowers attacks this directly with processes proven under pressure.
Cross-platform expansion matters too. A developer using Cursor and another using Claude Code can share exactly the same Superpowers configuration in a project repository. Skills live in .agents/skills/ and are automatically discovered by all supported agents. The methodology travels with the codebase.
Resources
GitHub Repository (MIT)
Complete skills documentation
Release notes
Are you already using Superpowers with more than one agent? Which new skill proved most useful — git worktrees, parallel agents, or the security tools? Let us know in the comments.