The cloud is less dramatic than last week’s version, and more useful for that. The largest token is still “skill” (1,242), followed by “agent” (971) and “claude” (671). Those three words are not single-project accidents. They are the shared grammar of repos that want to be installed into, wrapped around, or remembered by coding agents.
The next layer shows how that grammar is splitting. “agents” (355), “codex” (202), “opencode” (103), and “harness” (127) are platform words. “headroom” (295), “codegraph” (175), “graphify” (151), and “knowledge” (104) are context words. One camp names the actor; the other names the budget and map the actor needs. The week’s vocabulary shift sits right there: repo names are moving from model identity to agent operating conditions.
“Open” (457) still matters, but it is no longer the whole story. It points to open-design and other open alternatives, yet it now sits below the agent words. The stronger prefix this week is almost invisible because it is a suffix: “skill.” It appears as a product category rather than a decorative label. When skill outranks agent, the market is telling you that behavior patches are easier to package than full applications.
Proper names remain, but they read differently. “hermes” (307), “mimo” (278), “cl4r1t4s” (264), “odysseus” (235), and “turbovec” (137) are not a single mythological wave. They are project-name tokens that became large because their repos rose. That is useful, but it should not be over-read as culture. The more stable signal sits in reusable nouns: skill, agent, Claude, open, design, desktop, knowledge.
The surprising second ring is document and workspace language. “markitdown” (195), “ppt” (149), “desktop” (119), “notebook” (106), and “knowledge” (104) suggest a pipeline around files: convert them, organize them, then give them to an agent. That matches the repo list. MarkItDown turns messy inputs into Markdown; Tolaria turns Markdown folders into a local knowledge base; multiple graph tools turn folders into queryable maps.
The missing words are still useful. “gpt” and “llama” do not define the cloud; “llm” appears at 267, but it is generic. The names this week live one level above the model. They talk about agents, skills, harnesses, context, desktop surfaces, and documents. The model is assumed. The work is in making it behave.
This word cloud is tokenized from repository names, so it measures naming, not a curated taxonomy. A hot repo can inflate its own words. Even with that caveat, the trend is plain: GitHub projects are no longer naming themselves mainly after models. They are naming the control layer around models.