Superpowers Passes 290,000 GitHub Stars as Token Costs Draw Fire

Superpowers, an open-source methodology that walks coding agents through brainstorming, planning and reviewed implementation, had 290,504 stars and 25,995 forks on its GitHub repository as of Wednesday, and GitHub's own trending page showed it picking up another 528 stars over the day. The MIT-licensed project was created Oct. 9, 2025, according to the repository's own metadata, putting less than a year between its start and a star count most venture-backed startups never reach.

The framework's README describes a three-stage workflow: brainstorming that turns a request into a specification, planning that produces what the document calls an implementation plan clear enough for "an enthusiastic junior engineer" to follow, and subagent-driven development, in which separate agent instances write, test and review each step. Founder Jesse Vincent, who built Superpowers at Prime Radiant, the applied-AI lab he founded, lists more than 16 supported coding agents in the README, including Claude Code, Cursor, Codex, Gemini CLI, OpenCode and Devin CLI. The project's most recent release, v6.4.1 on Sept. 19, added a new diagnosing-superpowers skill meant to read a failed session's transcript and report what went wrong.

The release history behind that shows a project iterating under its own load. Version 6.0.0, shipped June 16, rewrote the subagent-driven development review process to use one reviewer per task instead of two. Version 6.1.0, on June 30, compressed what the project calls its bootstrap sequence specifically to cut per-session token costs, the same category of complaint now filling issue 743. Version 6.3.0, on Aug. 12, added three more supported harnesses, Devin CLI, Hermes Agent and Grok Build CLI, and changed the brainstorming stage to scale its questions to the size of the task rather than running the same ceremony on every request.

What the issue tracker shows

Superpowers had 398 open issues as of Wednesday, and several of the most-discussed describe the review process consuming far more of a coding agent's token allowance than working without it. In issue 1152, a user running Superpowers 5.0.7 against Codex 0.120.0 wrote that a 20-step implementation plan exhausted their entire five-hour ChatGPT Plus token allotment in 20 to 30 minutes, completing 17 of 20 steps before stopping. A second user, posting as aidada, reported the same exhaustion on a Pro-tier account and wrote that Codex's subagents were automatically downgraded from "5.4 medium" to the smaller "5.4-mini medium" once the account's remaining budget fell below roughly 30 percent.

A separate issue, 743, collected reports of the framework slowing everyday work to a crawl: user OZRay888 wrote that requests which normally finished in under 10 minutes were taking 30 to 45 minutes for the same code, and user latinrev described a run of close to two hours for an implementation that used to take 30 minutes at most. User kvan wrote that a single review sub-agent would "spin through dozens of calls to sed, grep, hexdump, git logs etc." to check a step that only wrote one short file, calling it a review process that runs at full intensity regardless of the size of what it is reviewing. User mgsgde wrote in the same thread that they had uninstalled and reinstalled the framework twice, adding that the brainstorming stage was worth keeping but that implementation had become "simply too slow," to the point that using a model directly produced the same quality of output in less time.

Report	Environment	What happened
Issue 1152	Codex 0.120.0, ChatGPT Plus	5-hour token budget exhausted in 20-30 minutes on a 20-step plan
Issue 1152 (aidada)	Codex, ChatGPT Pro	Subagents downgraded to 5.4-mini once budget fell below 30 percent
Issue 743 (mttetc)	Claude Code	Token use reported at 3x to 5x normal
Issue 743 (OZRay888)	Codex	Tasks that took 10 minutes began taking 30 to 45 minutes

What Vincent says, and what's still open

Asked in issue 743 what specifically was slow, Vincent replied that "review loops are a standard part of the superpowers process," adding that OpenAI's own token limits have been shifting and that a weaker model such as GPT-5.4 mini can make the review process take longer by producing work that itself needs correction. He wrote that without seeing the actual planning documents involved, he had no way to judge what was going on in any individual report.

The README carries its own disclosure: by default, a visual companion feature in the brainstorming stage loads a Prime Radiant logo from the company's website, which the document says is used to track how many sessions are running which version, a setting a user can turn off with the SUPERPOWERS_DISABLE_TELEMETRY environment variable. Separately, the project states plainly that it does "not generally accept contributions of new skills," limiting outside review to changes on skills the maintainers already ship.

Superpowers is not the only agent framework climbing GitHub's trending chart this week. Google's Agent Executor passed 7,600 stars days after its own v0.3.0 release, part of the same wave of agent-orchestration tools competing for the same developers now weighing Superpowers against its token bill. The same week, OpenAI dropped a cache-key requirement from GPT-6, a reminder that the token-pricing rules underneath a framework like Superpowers shift on a schedule the framework itself does not control. Vincent has already cut per-session costs once, in the v6.1.0 release, compressing what the changelog calls the framework's bootstrap sequence. Whether a similar change reaches the review loop is the variable that will decide if the star count keeps outrunning the complaints, or the other way around.