title: "GitHub Trending Today: mattpocock/skills Doubles Its Velocity and free-claude-code Refuses to Decelerate" description: "James Chen on today's GitHub trending — mattpocock/skills pulled 5,645 stars in a single day more than doubling yesterday's number, free-claude-code re-accelerated to 2,949 stars defying its own deceleration narrative, Microsoft dropped VibeVoice into the open, and TradingAgents is sitting at 54,000 stars which should prompt more questions than it usually does." publishedAt: "2026-04-28" author: "James Chen" category: "open-source" tags: ["github-trending", "open-source", "ai-tools", "claude-code", "agent-skills", "llm", "developer-tools", "voice-ai", "trading-agents", "memory-layer"]

Something strange happened to mattpocock/skills between yesterday and today.

Yesterday it pulled 2,519 stars. I covered it, called it the lead story, described the skills format and the distribution mechanism and why it was the right thing at the right moment for engineers who'd been running Claude Code long enough to be frustrated by re-prompting the same planning sequences from scratch.

Today it pulled 5,645.

That is not a typo. In a single day, the star velocity more than doubled. The repo now sits at roughly 32,000 total stars, which puts it in the upper tier of actively-maintained developer tools by sheer number — though the number is less interesting to me than the trajectory. 5,645 stars in one day on a repo that already had a strong prior day means this hit a second, different audience. The first 2,500 came from people who already knew about Claude Code agent skills. The next 5,600 came from people who had no idea this format existed and found it interesting enough to star without necessarily understanding the full context yet.

That second wave is the one worth watching. The first wave is enthusiasts. The second wave is potential adopters.

mattpocock/skills: Still the Most Interesting Thing on the Board

mattpocock/skills has not changed materially since yesterday. The skills are the same — to-prd, to-issues, grill-me on the planning side; tdd, triage-issue on the development side. The distribution mechanism is the same — npx skills@latest add mattpocock/skills/<skill-name> drops the skill into your project configuration. Matt Pocock, who runs the Total TypeScript newsletter and has the kind of audience that actually reads and implements the things he publishes, extracted this from his personal .claude directory.

What changed is the awareness surface. It is being shared in circles that do not primarily identify as Claude Code users.

That is structurally interesting. Early adopters of agent skills frameworks are, by definition, the people most willing to work against rough edges — they read the docs, file the bugs, figure out the workarounds. The people arriving from the second wave often aren't. They install a skill, hit friction, and leave. The ones who stay generate the bug reports that make the framework actually usable for everyone else. Both groups are necessary, but they are doing different jobs.

What I'd tell someone encountering this for the first time today: the installation mechanism works cleanly and the individual skills are documented with enough specificity to be useful. Start with grill-me if you have a tendency to start writing code before you've resolved the key design ambiguities. It will annoy you, which is the point. The to-issues skill is worth running once you have a plan — it slices work into independently-completable units using vertical slices rather than sequential phases, which means you can pick up a task mid-session without re-reading a plan document to figure out where things stand.

The thing I'm watching now is whether Pocock publishes new skills as he develops them, or whether this repo is a one-time export of his current .claude directory that will drift as he iterates on his private setup. Both outcomes are plausible. The commit history will tell the story over the next few weeks.

Worth cloning? If you are using Claude Code daily and you keep re-establishing the same planning rituals at the start of every session, yes. The grill-me skill alone is worth the two-minute install if you've ever shipped something that had an obvious flaw you would have caught if someone had pushed back on your assumptions before you wrote a line of code.

free-claude-code: The Repo That Will Not Cooperate With My Predictions

Yesterday I wrote that Alishahryar1/free-claude-code was "settling around 1,000 to 1,500 as it gets picked up by second-tier referrers." It did 1,701 yesterday. I was pleased with that call.

Today it pulled 2,949.

The total is now 16,583 stars. Something shared it again — I don't know what, the star growth curve doesn't carry source attribution — but the growth pattern looks like another major newsletter or social post rather than organic search discovery. Organic search doesn't do 2,949 stars in a day on a repo with no new commits.

The framing of the repo — "use claude-code for free in the terminal, VSCode extension or via discord" — is accurate about the mechanism but somewhat optimistic about the experience. You are not getting the full Claude Code experience when you route through a proxy. You are getting the Claude Code interface attached to whatever the proxy is pointing at, which may be NVIDIA NIM's free tier, OpenRouter, DeepSeek, or a local inference backend, depending on configuration.

The model quality matters more in agentic loops than it does in single-turn chat. An agent making a sequence of decisions — read this file, determine what needs to change, edit it, run the tests, interpret the results — needs to maintain coherent context across many tool calls. Smaller models drift. They re-run operations they just ran. They write code that doesn't match the style of the surrounding context. They hallucinate file paths they haven't seen. The Claude Code interface is doing everything right. The intelligence behind it is doing its best given what you've pointed it at.

That said: 16,000 people have starred this repo, and some meaningful fraction are using it successfully, otherwise the growth would have flattened out faster. "Not identical to the paid product" is not the same as "not useful." If your constraint is the subscription cost and your use cases are contained — single-file edits, bounded refactors, questions about code you don't understand — this is worth trying. Just calibrate your expectations to the model you're actually running, not to the interface wrapping it.

microsoft/VibeVoice: Open-Source Voice AI With a Marketing Name

microsoft/VibeVoice picked up 757 stars today against a total of 43,510. The description is "Open-Source Frontier Voice AI," which is four words doing a lot of work in different directions.

The repo is Microsoft's open-source implementation of a frontier voice AI stack — real-time speech synthesis and recognition intended to push the boundary on latency and naturalness rather than to expose a commercial product. The "frontier" qualifier is specific: this is not a wrapper around Azure Cognitive Services. It's research-grade code that Microsoft is publishing to let the community work with, extend, and benchmark against it.

43,510 total stars is substantial for a Microsoft research project. Most Microsoft AI open-source repos accumulate stars slowly over years as developers integrate them or cite them in papers. That many stars suggests this has either been building for a long time or it solved a problem resonant enough to get shared widely early. Given that the "today" count of 757 is roughly consistent with a repo that hit a trending moment and is now sustaining a normal discovery rate, I'd guess the total has been building over months rather than exploding recently.

The "Vibe" in VibeVoice is, I assume, a nod to the vibe coding moment, which has become shorthand for AI-assisted development to the point where Microsoft's marketing team apparently felt comfortable attaching it to a voice synthesis project. The name implies something casual and fast. The underlying project, based on what I can read of the structure, implies something more like serious infrastructure for people who need voice AI that can handle the edge cases — accents, background noise, domain-specific vocabulary — rather than a polished demo that falls apart when you deviate from the training distribution.

If you are building anything that involves voice interaction, this is worth looking at seriously. The test I'd apply before committing to any frontier voice project: check the latency at the 95th percentile under real network conditions, not the median under ideal lab conditions; check whether the synthesis quality is actually distinguishable from commercial alternatives in a blind test on your specific use case; and look for community benchmarks rather than author-reported numbers. At 43k stars, those benchmarks probably exist. Find them before you make a build-vs-buy decision.

gastownhall/beads: Sustained Relevance Is Its Own Signal

gastownhall/beads added 498 stars today, bringing its total to 22,354. This is at minimum the third consecutive day I've seen it trending in the AI and developer tools space.

I covered beads in detail in yesterday's piece, so I won't repeat the full analysis. The short version: it's a distributed graph issue tracker for AI agents built on top of Dolt, which is a SQL database with full git semantics. Tasks in beads have hash-based IDs to prevent merge collisions, explicit dependency links, and a compaction mechanism that summarizes closed tasks when they start consuming context window. There's an MCP server so your agent can query and update the task graph through tool calls rather than reading a markdown file.

What the sustained trending position tells you is that the word is spreading through a community with a concrete, shared problem: they are running multi-session agent workflows and markdown todo files are failing them. This is not a theoretical concern. Anyone who has tried to use Claude Code on a task that spans more than two sessions knows the re-context problem — you spend five minutes at the start of every session narrating the history to your agent, because the only shared memory between sessions is whatever you put in a text file.

Beads is a structural solution to that problem rather than a prompt engineering solution. The distinction matters because prompt engineering solutions require you to do the work — write the summary, maintain the file, paste it in at the start of each session. Structural solutions run that work through infrastructure that doesn't depend on you remembering to do it.

498 stars today without a new feature launch or a viral social post means people are continuing to find this organically, which is the kind of growth that tends to compound. If I had to pick one thing from this week's trending for someone who hasn't tried it yet, this is still the recommendation.

ComposioHQ/awesome-codex-skills: Riding the Wave, Legitimately

ComposioHQ/awesome-codex-skills added 638 stars today against 3,120 total. It is a curated list of practical Codex skills for automating workflows across the Codex CLI and API — the OpenAI Codex equivalent of what mattpocock/skills represents for Claude Code.

The velocity correlation with mattpocock/skills is not an accident. When a skills-format repo goes viral in one ecosystem, it surfaces adjacent repos through search and recommendation systems. Some fraction of the people discovering agent skills for the first time via Matt's repo are landing on this one because it comes up alongside it. 638 stars in one day from a list repo is a number that requires an external traffic source; the organic discovery rate for awesome lists is much lower than that.

The list itself covers GitHub integration, code review automation, test generation, documentation workflows, and a range of other common developer use cases. The quality variance is higher than in a single-author repo — that is the nature of community-curated lists — but the coverage is broad enough to be useful as a starting point if you're using Codex CLI and you find yourself re-prompting the same sequences repeatedly.

Matt's repo is his actual daily workflow. This repo is a community aggregate of what many people think is worth sharing. Both have value. They're solving slightly different discovery problems.

TauricResearch/TradingAgents: 54,000 Stars, Reasonable Skepticism Required

TauricResearch/TradingAgents pulled 248 stars today against a total of 54,088. The description: "TradingAgents: Multi-Agents LLM Financial Trading Framework."

54,088 total stars is a significant number. For reference, that places it alongside well-established projects like FastAPI. The star count reflects the appeal of the underlying idea: a multi-agent LLM system where specialized agents handle different aspects of financial analysis — one reads sentiment, one analyzes technicals, one tracks fundamentals, a coordinator synthesizes their outputs into a position recommendation — and collectively attempt to make better trading decisions than a single model or a single analyst.

The idea is obviously appealing. Every retail investor who has read about AI wants to believe there's a framework that can extract alpha from markets if it just coordinates enough agents with enough data. The 54,000 stars are the record of that belief.

The question those stars do not answer is whether the agents' signals are actually predictive of future price movements in a way that survives realistic transaction costs and slippage. This is the question that matters if you are considering using this for anything beyond a research experiment, and it is also the question most conspicuously absent from project landing pages in this category. Backtesting against historical data under idealized conditions is easy to do and easy to make look impressive. Building a trading system that remains profitable in live markets after you account for the spread, the commission, the market impact of your own orders, and the fact that every other quant in the world has access to the same historical data you're training on — that is a different problem entirely.

This is, correctly, described as a research framework rather than a production trading system. It is a platform for experimenting with multi-agent approaches to market analysis, and as an experimental platform it has genuine value. You can learn things about agent coordination, tool call sequencing, and multi-source information synthesis by running this that are applicable outside the trading domain.

What I would not do: run this against a live brokerage account without extensive paper trading, realistic cost modeling, and a clear-eyed read of the actual signals it's generating versus what it would need to generate to be profitable. The 54,000 stars will not protect your capital.

The theme across today's trending board is consolidation. The agent skills format that was experimental three weeks ago now has a distribution mechanism, a star velocity story, two separate communities building lists around it (Claude and Codex), and a second day of viral growth on the originating repo. The memory layer problem that was theoretical six months ago now has an open-source Go implementation with 22,000 stars and enough sustained traction to appear on the trending board for a week straight. Microsoft is open-sourcing voice AI infrastructure that it would have shipped as an API-gated Azure service two years ago.

The agents themselves are not changing that fast. The scaffolding around them — skills packages, memory layers, evaluation infrastructure, open model stacks — is changing fast. That scaffolding is what determines whether an agent is useful across sessions, across projects, and across people who are not the person who configured it. Today's trending board is mostly that scaffolding. That is the right thing to be building right now.