GitHub Trending #1: what I'd actually clone (and the one I wouldn't)
I ran /compact on my own working session about an hour before I sat down to write this — squeezing a long, sprawling conversation back into a context window so the agent could keep going. So it was a small, silly joy to open this week's GitHub Trending list and find a repo that exists to do exactly that job, better. That's the thing about Trending: buried in the noise there's usually one project that reads your mind.
The noise is the problem, though. GitHub Trending has become a minefield of star-farmed shells — grand names, nine commits, a README that's eighty percent animated GIF. So I've started doing the boring thing: actually opening the repos as they climb, reading past the marketing, and keeping only the few that survive the read. This is the first of those write-ups. The good news this week is that the crop was unusually clean. Here's what I'd actually clone, grouped by the itch it scratches — and the one I'd walk straight past. If it earns its keep, it'll become a regular thing.
Everyone is building the scaffolding around agents
The clearest signal this week is that the interesting work has moved up a layer. Not "here's another coding agent," but "here's the plumbing to run agents without them tripping over each other."
tutti is a shared workspace that sits over Claude Code and Codex so several agent sessions can @-mention one another, pass files and outputs around, and delegate sub-tasks — instead of each one re-deriving the same context in its own silo. TypeScript and Go, Apache-2.0, a couple of thousand stars and a real commit history behind it. If you run more than one agent at a time — and I increasingly do — this is the shape of the tooling that's coming.
shepherd is the one I found quietly ambitious. It wraps an agent's execution in a reversible, Git-like trace: fork a run, replay it, revert it, with syscall-level sandboxing and per-repo permission grants. That is almost exactly the governance layer I complained was missing when I wired a coding agent into a live Drupal site over MCP. It's alpha and it's Python, but it's the right idea pointed in the right direction.
Two smaller ones in the same spirit. watch-skill gives an agent the ability to genuinely watch video — 1,800-odd sources via yt-dlp, frames plus OCR plus Whisper transcripts, indexed and exposed over MCP. That closes an awkward gap; I've wanted to hand an agent a conference talk more than once. And compact-plus — my /compact saviour — is a Claude Code plugin that backs up the transcript, writes a ten-section state summary before compaction, and re-injects recovery guidance after. A shell script that does one annoying thing well.
If you'd rather understand agents than orchestrate them, tau from Hugging Face is the pick: a small, deliberately readable terminal coding agent built as a reference example — reads files, edits, runs commands, streams output, across OpenAI, Anthropic and OpenRouter. It's the "read the source to learn how the trick actually works" repo of the week.
And if you want somewhere safe to let them loose, boring-computers spins up on-demand Firecracker microVMs — desktop-over-VNC or headless shell — that an agent can drive, with roughly 35ms forking and TTL self-destruct. Self-hostable, which on this blog is always the part I care about.
People are rewriting the runtime from scratch (again)
The other loud theme this week: hand-built language runtimes, in Rust and C, apparently for the pure hell of it — and they are not toys.
colibri is my favourite object of the week precisely because it's honest about being impractical. It runs a 744-billion-parameter mixture-of-experts model on about 25GB of RAM by streaming its 21,504 routed experts off disk instead of holding them in memory — in roughly 2,400 lines of pure C. It manages a blistering ~0.05 tokens per second, and the README says so, out loud, without flinching. That's exactly my kind of repo: a beautiful idea, benchmarked honestly, that teaches you something even though you'd never ship it.
More plausibly useful: pon is a JIT/AoT native compiler and runtime for Python 3.14 written in Rust with no interpreter and no bytecode, byte-exact differentially tested against CPython. And lumen is a JavaScript engine in Rust with three execution tiers — tree-walker, bytecode VM, ARM64 JIT — passing 100% of test262. Both are the "surely this is a weekend toy… oh, it passes the conformance suite" kind of project that leaves you inspired and lazy in equal measure.
…and two that just fix something today
knockoff is the most immediately useful thing on the list: a browser extension that detects and dims the trademark-squatting pseudo-brand listings that infest Amazon, while leaving genuine brands alone. Multi-browser, curated brand datasets, ~1.8k stars. I installed it before I'd finished reading the README.
Honourable mention for the Mac crowd: davit, a native SwiftUI GUI for Apple's own container platform that talks to it over XPC rather than shelling out to the CLI — a real gap for anyone on Apple silicon who wants an OrbStack-style experience.
The one I'd walk past
Every trending list has its tell. This week it's motion-anything: 389 stars sitting on nine commits, a README wallpapered in GIFs promising "403 recipes, 77 effects, 2,680 icons" and a "chat-native motion engine." That ratio — lots of stars, almost no history, heavy marketing — is the classic star-farm signature. Maybe it grows into the claims. Today it's a screenshot with a landing page attached — and knowing how to spot the pattern is what keeps the genuinely good work above from getting drowned out.
Links
- tutti — multi-agent workspace
- shepherd — reversible agent runtime
- watch-skill — video for agents over MCP
- compact-plus — better /compact for Claude Code
- tau — readable reference coding agent
- boring-computers — Firecracker microVMs for agents
- colibri — disk-streaming MoE inference in C
- pon — native Python compiler in Rust
- lumen — JavaScript engine in Rust
- knockoff — hide Amazon knock-off brands
- davit — SwiftUI GUI for Apple containers