title: "What Actually Trended on GitHub Today (April 30, 2026)" description: "Jesse Vincent built a keyboard. Now he has 173,000 stars on an agentic development methodology. Warp opened an issues repo yesterday and immediately became the day's top trending project. Heretic does something technically fascinating to LLMs that will make some people uncomfortable. Here's the real breakdown." publishedAt: "2026-04-30" author: "James Chen" category: "open-source" tags: ["github-trending", "ai-tools", "open-source", "llm-agents", "developer-tools", "claude-code", "agentic-dev"]

Jesse Vincent — known on GitHub as obra — has shipped open source hardware, firmware for mechanical keyboards, and more tooling for Perl than anyone should probably admit to. He is not, by background, someone you'd expect to see sitting at the top of GitHub trending with 173,000 stars on a software development methodology. And yet.

obra/superpowers has accumulated 173,686 total stars since it appeared in late March, picked up another 1,653 today, and is available as a plugin in the official Claude Code marketplace. The velocity on this repo is one of the more striking things I've seen this year — not because it's impossible to explain, but because the explanation reveals something specific about where developer tooling is right now.

What Superpowers actually is: a complete methodology for coding agents, layered on top of a set of composable skills. The summary version is that when you fire it up, your coding agent doesn't just start writing code at you. It asks what you're actually trying to build. It helps you write a spec. It shows you that spec in digestible chunks. It produces an implementation plan explicit enough that — as the README puts it with a particular kind of precision — "an enthusiastic junior engineer with poor taste, no judgement, no project context, and an aversion to testing" could follow it. Then, once you've signed off, it runs subagent-driven development: actual sub-agents working through engineering tasks, being inspected and reviewed, continuing forward.

The result, according to the README, is that Claude can work autonomously for a couple of hours at a time without deviating from the plan. I've seen similar claims before and most of them dissolve the moment you try them on a nontrivial project. Superpowers is different enough in its specificity that I'm inclined to believe it works at least as well as advertised — the key innovation isn't any single clever prompt, it's the structure. The spec-first, plan-second, execute-third workflow isn't novel; it's how good engineering teams already operate. Superpowers just encodes that workflow in a form an agent can follow automatically.

The installation is also notably non-painful. If you're on Claude Code, it's a single command: /plugin install superpowers@claude-plugins-official. It also works on Codex CLI, the Codex app, Cursor, VS Code Copilot, and a few others. The fact that Jesse built this to work across multiple agent platforms rather than betting on one horse is either pragmatic product sense or a sign that he doesn't trust any of them to dominate long-term. Given his history, probably both.

Worth cloning? If you're using Claude Code seriously for anything more complex than one-shot code generation, install this and run it for a week before deciding whether it changes how you work. The worst case is that you learn what kinds of structure actually help coding agents stay on track, which is useful knowledge regardless.

Yesterday, warpdotdev/warp created a new GitHub repository. It's an issues-only repo — the README explicitly says so — for bug reports, feature requests, and feedback. Today it picked up 12,822 stars, which put it at the top of GitHub trending by a significant margin. For reference, that's more stars in a single day than most serious open source projects accumulate in a year.

I want to say something honest about this number, because it would be easy to misread it.

Warp is a terminal. More precisely, it's a modern terminal that's been rebuilt around the concept of agentic development — it has its own built-in agent called Oz, it supports running Claude Code, Codex, and Gemini CLI inside it, and it's positioning itself as an "orchestration platform for cloud agents" rather than just a place where you type commands. The Warp team has been building this for a while. They have a real product. The issues repo opened yesterday because they're beginning to move parts of their codebase into the open.

The 12,822 stars aren't 12,822 new Warp users who discovered the terminal today. They're 12,822 developers who heard "Warp is going open source" and hit the star button, which is understandable behavior given what Warp has built. The Rust UI framework is apparently coming first, then potentially the client codebase. The server portion is staying closed.

What actually matters here: Warp going open source — even partially — changes the landscape for terminal tooling in a real way. The terminal is where most serious agent workflows actually run. Claude Code lives in it. Codex CLI lives in it. If the Warp UI layer becomes something people can build on top of, the downstream applications are genuinely interesting.

The star count is a sentiment metric, not a usage metric. But when 12,822 developers express that particular sentiment in one day, it tells you that people have been waiting for this, which is worth knowing.

TauricResearch/TradingAgents was in yesterday's list, had a strong day on Python trending again today, and at 56,258 total stars it's been one of the more sustained trending stories of the past week. The repo is 5 days old. That trajectory is worth pausing on.

The framework deploys specialized LLM-powered agents that mirror how actual trading firms structure their research teams: a fundamentals analyst evaluating company financials, a sentiment analyst parsing social media and news feeds, a technical analyst running indicators, and risk management on top. These agents don't just produce outputs — they engage in structured debates, with bullish and bearish researchers critically assessing each other's positions before anything gets handed to the trader agent.

v0.2.4, which landed this week, added structured-output agents for the Research Manager, Trader, and Portfolio Manager roles; LangGraph checkpoint resume (so you can pick up a long-running analysis run if it gets interrupted); a persistent decision log; and support for DeepSeek, Qwen, GLM, Azure, Claude 4.6, and GPT-5.4 as backbone models. The Windows UTF-8 encoding fix is also in there, which tells you the team is actually doing maintenance rather than just adding features.

My consistent position on this category of repo: the multi-agent architecture is legitimately interesting and worth studying. The framework is well-built. The research paper (arXiv:2412.20138) is real. None of that means you should trust the output of LLM trading agents with money you care about. The research on whether these systems beat the market consistently is preliminary, and "consistently" is doing a lot of work in that sentence. The demos look compelling because the demos are run on periods where the strategy worked.

What this is actually good for: learning how to build multi-agent pipelines where specialized LLMs hand off context across long workflows, run parallel analysis, and maintain state through a persistent log. That architecture pattern is broadly applicable. The LangGraph integration is well-implemented and worth understanding even if you're building something entirely different.

p-e-w/heretic had 69 stars today — modest by the standards of everything else in this list — but 20,235 total stars and a #1 repository of the day badge from Trendshift at some point in its history. It was created in February. I've been watching it for a while and I want to describe it accurately, which requires some care.

Heretic is a tool that removes safety alignment from transformer-based language models without post-training. The underlying technique is directional ablation — sometimes called "abliteration" in the ML community, documented in Arditi et al. 2024 — combined with a TPE-based parameter optimizer powered by Optuna. The differentiator Heretic offers over manual abliteration approaches: it runs completely automatically. You give it a model, it co-minimizes the number of refusals and the KL divergence from the original model, and it produces a decensored version that retains as much of the original model's capabilities as possible.

The benchmark table in the README is genuinely interesting. Against a baseline of 97/100 refusals for "harmful" prompts, both the Heretic-generated model and manually abliterated versions hit 3/100 refusals. But the KL divergence from the original model — which measures how much the decensoring damaged the model's general capabilities — is 0.16 for Heretic versus 0.45 and 1.04 for the manual approaches. Heretic, run unsupervised with default settings, produces a better result than human experts doing the same thing by hand.

That's technically impressive. The Optuna optimizer is doing something genuinely clever: treating abliteration parameter tuning as a hyperparameter optimization problem rather than a craft. Over 1,000 community-made Heretic models have been published to Hugging Face.

I'll state my practical take directly: this tool is going to be most useful to ML researchers studying alignment, red-teamers at AI companies, and people running local models for creative writing who find current safety guidelines overly restrictive. It has obvious misuse potential, and the repo's existence in a post where I'm recommending other repos shouldn't be read as a blanket endorsement. What I can say is that the technical approach — automated parameter optimization for post-hoc model modification — is a meaningful contribution to the literature on model editing, whatever you think about the application.

hugohe3/ppt-master appeared yesterday (created April 29), picked up 414 stars today, and has 9,749 total already. The premise: drop in a PDF, DOCX, URL, or Markdown file, get back a natively editable PowerPoint. Not images exported as slides. Not a web deck. Actual PPTX with real shapes, real text boxes, real charts, real animations encoded as OOXML.

The "how it works" is the interesting part: it's a skill (or workflow, in the Claude Code sense) that runs inside AI IDEs. You chat with the AI — "make a deck from this PDF" — and it follows the workflow to produce the PPTX on your machine. No upload to a server. No subscription. You pay for whatever AI model API calls the generation requires.

The demo in the README is a 12-page deck generated from a single WeChat article URL using Claude Opus 4.7. Full animations, per-element entrance effects, page transitions — all as real OOXML that plays natively in PowerPoint and Keynote. The README draws a useful contrast with other AI presentation tools: template fill-in tools, image-based tools (each slide is a picture), and HTML presentation tools. PPT Master does none of those things. Everything it outputs is directly editable.

The trade-off you're making here is setup time. You need Python, an AI IDE, and some patience with configuration. If you generate presentations occasionally, it might not be worth the setup. If you generate them regularly and you're paying $30/month for an AI presentation tool on top of your AI API costs, the math probably works in your favor pretty quickly.

Clone it? If you're doing regular document-to-presentation work, yes. The data locality argument alone — your files don't leave your machine except for the AI model call — is worth something in enterprise contexts where you're working with sensitive documents.

mattpocock/skills held at 7,280 stars today, making it the second day in a row it's been in the top positions on trending. If you read yesterday's roundup, you know the pitch: composable agent skills, slash commands and prompt templates designed for Claude Code and Codex, built around the principle that "don't start coding until you actually understand what we're building" is a virtue rather than a slowdown. The /grill-me pattern — forcing the agent to interrogate your requirements before writing anything — is the part people keep citing.

The fact that it's still at this velocity on day two is the interesting data point today. Repos that spike and immediately fall back are a specific kind of viral: they hit something algorithmically but the substance doesn't sustain interest. Skills holding its numbers means people are looking at it more carefully than they did yesterday, not less. That's a different kind of signal.

Stepping back from any individual repo: the shape of today's trending list is doing something consistent with what I've been watching for months. The top velocity repos are almost all about agent workflow structure — how to make coding agents stay on task, how to encode engineering discipline into a format a model can follow, how to split work between specialized sub-agents. The question shifted from "can agents write code" to "what constraints make agents actually useful on complex work." Superpowers, mattpocock/skills, even the Codex skills lists — they're all answers to the same underlying question.

Warp opening an issues repo is a physical embodiment of the same trend: the terminal is where agent workflows run, so the terminal is becoming an agent orchestration platform. That's not metaphorical. They literally built a cloud-agent orchestration product into a terminal application and then decided to open it up.

The one exception in today's list is Heretic, which is solving a different kind of problem — or at least a problem that starts from a very different set of assumptions about what language models should be allowed to say. Whether that's a problem worth solving is a question reasonable people disagree about. The technical execution is, regardless, quite good.

Tomorrow the slots will be different. The themes probably won't be.