This week did not have one runaway repo. It had a pile-up around the same idea: the fastest way to ship AI tooling is now a small layer around an agent someone already runs. last30days-skill led the board with +365 stars, followed by headroom (+295), apple/container (+283), agent-skills (+277), and MiMo-Code (+273). The top five alone split the week cleanly: agent skills, context compression, local containers, coding-agent practice, and a new code model project.
The “skill” cluster is still the center of gravity, and it has grown into a product category, with novel naming as its earliest form. last30days-skill packages a research workflow across Reddit, X, YouTube, HN, Polymarket, and the web. agent-skills is a production engineering skill set. taste-skill, mattpocock/skills, obra/superpowers, pm-skills, and andrej-karpathy-skills fill out the same pattern. The repo is often a Markdown file, a shell wrapper, or a small directory of rules. That smallness is the product feature: distribution is easy because the artifact is close to copy-paste size.
The second cluster is about shrinking the agent’s working set. headroom compresses tool outputs, logs, files, and RAG chunks before they hit the LLM. codegraph builds a local pre-indexed knowledge graph for Claude Code, Codex, Gemini, Cursor, OpenCode, Kiro, and Hermes Agent. graphify turns folders of code, schemas, scripts, docs, papers, images, or videos into a queryable graph. Understand Anything lands at #24 with a visual code graph for humans. These projects leave the agent in place and decide what it sees, in what shape, and at what cost.
Apple’s container at #3 is the week’s useful counterweight. It is not a prompt pack or an agent wrapper. It is infrastructure: Linux containers through lightweight virtual machines on Apple silicon. Its presence near the top matters because the agent-tooling story still needs a local execution substrate. When developers run coding agents, sandboxes, MCP servers, and generated projects on a laptop, boring isolation becomes part of the AI stack.
The rest of the board shows how wide that stack has become. cc-switch manages Claude Code, Codex, OpenCode, OpenClaw, Gemini CLI, and Hermes Agent from one desktop assistant. NousResearch/hermes-agent turns Hermes into a named agent. Panniantong/Agent-Reach gives agents eyes on Twitter, Reddit, YouTube, GitHub, Bilibili, and Xiaohongshu without API fees. ECC treats harness performance as a system of skills, instincts, memory, and security. The board is full of things that sit between a model and the user’s actual work.
There are two non-agent signals worth keeping. microsoft/markitdown at #12 is a reminder that conversion to Markdown has become a common pre-processing step for AI workflows. refactoringhq/tolaria at #23 is a desktop app for Markdown knowledge bases, with Git-backed vaults and AI-agent setup paths. In the same week, one repo turns files into Markdown and another turns Markdown folders into a navigable knowledge base. That is a clean pipeline: ingest the mess, normalize it, then let agents work against it.
The honest boundary is still attention versus adoption. These are rolling seven-day star deltas, and the word “skill” carries a lot of naming momentum in June 2026. Some of these repos may become durable tools; others are one-file experiments riding the current interface fashion. Still, the direction is hard to miss. The week of June 8 was not led by model repos. It was led by the thin layers that tell agents what to do, what to read, and how much context they are allowed to spend.