title: "GitHub Trending Today: A Financial Markets Model, Matt Pocock's Claude Skills, and Why free-claude-code Is Still Accelerating" description: "James Chen's daily read through GitHub trending — free-claude-code broke its own record at 4,007 stars today, mattpocock/skills is actually useful, Kronos treats candlestick charts like language, and DeepEP is very impressive if you happen to own 64 H800s." publishedAt: "2026-04-26" author: "James Chen" category: "open-source" tags: ["github-trending", "open-source", "ai-tools", "claude-code", "llm", "finance-ai", "computer-vision", "developer-tools"]

Yesterday I said free-claude-code was pulling 2,638 stars in a day and implied the velocity would flatten out. It pulled 4,007 today. The market is not listening to me about this, and I am updating my priors accordingly.

That repo leads the chart again, but there are four genuinely new things on today's list worth talking about — including a financial markets foundation model accepted at AAAI 2026, Matt Pocock's personal Claude skills directory, an agent memory layer that's doing something interesting, and Roboflow's real-time object detection model from ICLR. Let's get into it.

free-claude-code Is Still Accelerating, Which Tells You Something

Alishahryar1/free-claude-code cleared 4,007 stars on Sunday, up from 2,638 the day before. I covered the mechanics yesterday — a lightweight proxy that intercepts Claude Code's Anthropic API calls and reroutes them to NVIDIA NIM's free tier, OpenRouter, DeepSeek, or local models via Ollama, LM Studio, or llama.cpp. Two environment variables, no modifications to Claude Code itself.

The velocity increase is what I want to flag today. Repos usually peak on day one and decay. This one is doing the opposite. That pattern typically means it's getting shared by larger accounts after the initial discovery wave — developers with audiences who don't watch GitHub trending but do watch Twitter and YouTube. The implication is that a very large population of developers either can't justify Claude's subscription cost, wants to evaluate before committing, or actively prefers routing through their own infrastructure for data reasons.

The technical story is unchanged from yesterday: you get the Claude Code interface, you lose Claude's actual intelligence unless you're routing to a model that can hold its own in agentic loops. The heuristic tool parser that converts text-formatted JSON tool calls into structured tool use is the most practical piece of engineering in the repo — smaller models leak their tool calls as prose constantly, and catching that heuristically rather than blowing up is the difference between a janky demo and a usable session.

One addition worth noting from the README I missed yesterday: there's a Discord and Telegram bot feature that does remote autonomous coding with tree-based threading and session persistence. That's not just a hack — that's a multi-platform async coding interface. Whether you want unsupervised agents doing things on your behalf through a Discord bot is a risk tolerance question, not a technical one.

mattpocock/skills: The Claude Code Skills Directory That Actually Reflects Real Workflows

mattpocock/skills picked up 1,139 stars today. Matt Pocock is the person behind Total TypeScript, so he knows exactly what the failure modes of complex TypeScript codebases look like, and it shows in what he built.

The repo is his personal .claude directory, published as installable agent skills. The range is broader than I expected. On the planning side: to-prd synthesizes the current conversation into a Product Requirements Document and files it as a GitHub issue, to-issues breaks a plan into independently-grabbable vertical slices, grill-me relentlessly interviews you about a design decision until every decision branch is resolved. On the development side: tdd implements test-driven development with an explicit red-green-refactor loop, triage-issue explores the codebase to find a root cause and files a structured issue with a TDD-based fix plan.

design-an-interface is the one I'd pay attention to as a pattern. It runs parallel sub-agents to generate multiple radically different interface designs for a module. Most of us are trained to think of design as a linear process: gather requirements, design, implement. The parallel sub-agent approach forces breadth before commitment. As the cost of running parallel agent queries drops — and it's dropped substantially in the last twelve months — this becomes a real workflow tool rather than a research demo.

The install command is npx skills@latest add mattpocock/skills/<skill-name>. The fact that each skill is individually installable rather than an all-or-nothing package is good design. You can adopt the triage-issue flow without importing someone else's opinions about how to write PRDs.

Worth forking, worth customizing. The value isn't in copying his exact skills verbatim — it's in understanding the structure well enough to codify your own team's conventions.

Kronos: A Foundation Model That Learned to Speak Candlestick

shiyu-coder/Kronos showed up at 283 stars today and caught me off guard. It was accepted at AAAI 2026, and the description — "A Foundation Model for the Language of Financial Markets" — is either a bold claim or a precise one, depending on what they actually built.

It's a precise one. Kronos is a family of decoder-only transformers pre-trained specifically on K-line (candlestick) sequences from 45 global exchanges. The architecture has two components: a domain-specific tokenizer that quantizes continuous OHLCV data (Open, High, Low, Close, Volume) into hierarchical discrete tokens, and then a large autoregressive transformer trained on those tokens. They didn't adapt a language model to handle numbers. They built a tokenization scheme specific to financial time series and then trained from scratch.

The model zoo goes from Kronos-mini at 4.1M parameters to Kronos-base at 102.3M, all on HuggingFace. The large variant at 499M is closed for now. Fine-tuning scripts have been available since August 2025. There's a live BTC/USDT forecast demo, which I checked — the predictions have plausible shape but I wouldn't bet the portfolio on them, and neither should you.

What I find more compelling than any specific forecast is the architectural pattern. A domain-specific discrete tokenizer feeding an autoregressive transformer is a generalizable approach. Infrastructure metrics, sensor readings, genomic sequences, any high-noise time series where continuous numeric representations don't capture the right structure — the same two-stage approach applies. The Kronos paper is worth reading for the tokenizer design alone, regardless of whether you care about financial markets.

The standard disclaimer: if this model maintained a reliable edge, it wouldn't be on GitHub. Open-sourcing a profitable signal kills the signal. But as a research starting point for financial time series work, and as a case study in domain-specific tokenization, it's one of the more interesting things I've seen trend this week.

deepseek-ai/DeepEP: Impressive in Proportion to Your GPU Count

deepseek-ai/DeepEP came in at 189 stars today. DeepSeek keeps releasing the actual infrastructure behind their models, and that's valuable at a different level than the model weights themselves.

DeepEP is a communication library for Mixture-of-Experts models — the all-to-all GPU kernels (dispatch and combine) that bottleneck expert-parallel training at scale. The benchmarks: 153 GB/s dispatch bandwidth over NVLink intranode, 58 GB/s over RDMA at 32 nodes, FP8 support, a hook-based communication-computation overlap method that doesn't consume streaming multiprocessors during the overlap. Tencent's network team contributed optimizations last week that pushed performance up 30%.

Those numbers are real. The test environment — H800 nodes with CX7 InfiniBand at 400 Gb/s — is where those numbers exist. If you have H800s connected to 400 Gb/s InfiniBand, you already have a team that knows what to do with this library. If you don't, the code is useful reading for the overlap technique and the RDMA optimization patterns, which will filter into other frameworks eventually.

Worth following for the implementation ideas. Worth cloning if you're actually running MoE training at scale. Otherwise: read the paper, bookmark the hook-based overlap approach, move on.

MemoriLabs/Memori: Agent Memory That Works From Behavior, Not Just Transcripts

MemoriLabs/Memori pulled 124 stars today, and the positioning is interesting: "Memory from what agents do, not just what they say."

Most agent memory systems store message history and retrieve relevant chunks via similarity search at the start of each session. That works for Q&A recall. It breaks down when your agent is executing multi-step tasks across sessions and needs to remember intermediate state — what it decided to do at step 3 and why, what failed at step 7, what the user's unstated constraints turned out to be over multiple interactions. Transcript retrieval doesn't capture any of that cleanly.

Memori's approach wraps your existing OpenAI or Anthropic client with two lines of SDK code. Conversations persist and recall automatically. The attribution API assigns memory to specific entity/process pairs — mem.attribution(entity_id="user_123", process_id="support_agent") — which means in a multi-agent system you can track what each agent knows about each user separately. That's the right primitive for production agent architectures.

The caveats: it's a cloud service, not a fully self-hosted library. The SDK is Apache-licensed; the storage infrastructure is managed. For any project where conversation data can't leave your infrastructure, this doesn't work as-is. For commercial agent applications where data requirements permit a third-party memory service, the operational simplicity is real — you don't have to build and maintain a vector store, embeddings pipeline, and retrieval system.

Whether it solves the hard version of the agent memory problem — tracking what an agent did and not just what it said — isn't clear from the README. But the attribution API and the "behavior-based" framing suggest someone has thought about this more carefully than most. Worth a closer look if you're building stateful agents.

ComposioHQ/awesome-codex-skills: Curated Automation for the Codex CLI

ComposioHQ/awesome-codex-skills hit 188 stars today. It's a curated list of practical skills for the Codex CLI and API — automating workflows across GitHub, Notion, Linear, Slack, and similar surfaces. Composio makes tooling for connecting agents to APIs, so this is partly a product showcase and partly a genuine community resource.

The "awesome list" format lives or dies by curation quality. What I'd look for before recommending it: are the listed skills battle-tested on real workflows, or are they demos? An awesome list full of demos is just noise with better formatting. Worth a quick scan if you're building Codex-based automation. Don't assume everything listed has been stress-tested.

roboflow/rf-detr: ICLR 2026, DINOv2 Backbone, Benchmarked on Real Hardware

roboflow/rf-detr is today's sleeper at 59 stars — modest velocity, but the underlying work is solid. It's a real-time object detection and instance segmentation model from ICLR 2026, built on a DINOv2 vision transformer backbone, SOTA on Microsoft COCO.

What separates this from typical conference paper repos is the Roboflow engineering team's habit of actually shipping things. The install is pip install rfdetr. The API is clean. Benchmarks are on a T4 GPU, which is the kind of hardware real inference workloads run on rather than the A100s that paper figures use. Detection and segmentation come out of the same model via one consistent API. Apache 2.0 for the base models, commercial license for the XL and 2XL variants — a sustainable structure.

If you're running YOLOv8 or RT-DETR on something and you care about accuracy, this is worth benchmarking on your task. The DINOv2 backbone transfers well across domains. The instance segmentation addition without splitting into a separate model is practically useful — needing both outputs is common, and running two separate models is wasteful.

This is the kind of repo where the GitHub velocity understates the value. 59 stars today because it's not a jailbreak and it doesn't have "free" in the title. Clone it, run your benchmark task, compare numbers. If it holds up, it's a real upgrade.

The pattern I keep seeing in today's list: the repos with the highest velocity are the ones solving billing friction (free-claude-code, up 50% day-over-day), and the repos with the highest technical value are pulling 50-300 stars. That gap is not new, but it's getting wider as the AI tooling market matures and the difference between "goes viral" and "actually useful" becomes easier to see.

My short list for today: if you do computer vision work, run rf-detr on your benchmark task. If you're using Claude Code, spend twenty minutes with mattpocock/skills and steal what fits your workflow. If you do financial time series work for any reason, read the Kronos paper. If you're building stateful multi-agent systems, Memori's attribution API is worth understanding even if you build the memory layer yourself.

Everything else on today's list: interesting to read, harder to use in practice than the star counts imply. The star is not the review.