The model isn't the moat anymore; the edge is what you load into it.

`openai/skills` is OpenAI's official catalog of Agent Skills for Codex — folders of instructions and scripts in three tiers (`.system` auto-installed, `.curated` installable via a `$skill-installer` tool, `.experimental`). It's at 22.5k stars. The news isn't the skills inside; it's that OpenAI adopted the same SKILL.md open format Anthropic released, so a skill written for Claude Code drops into Codex unchanged. Reach for it when you maintain a procedure — a deploy checklist, a house code style, a data pipeline — and want it to ride along whichever agent you're driving that day. Adopting this deletes the per-agent prompt fork: the AGENTS.md you wrote for Codex and the CLAUDE.md you wrote for Claude, kept in sync by hand. Tradeoff: the repo ships with no license, an odd omission for something OpenAI wants you to copy from. Read the skills for technique; don't assume you can redistribute them.

Garry Tan's `gbrain` (MIT, 23k stars) is a knowledge and synthesis layer an agent installs and operates itself over MCP — hybrid vector-plus-keyword search, an auto-linked knowledge graph, durable subagent job queues, an eval harness, and 43 curated skills. It runs with Claude Code, Codex, OpenClaw, or Hermes. Use it when you want an agent with persistent, cited memory of your own docs and notes instead of re-pasting context every session — it synthesizes answers with sources and gap analysis rather than dumping raw pages. It replaces the pile of ad-hoc RAG scripts and the "paste the spec again" ritual at the start of every chat. Tradeoff: setup is your agent running roughly 30 minutes of install steps against a local database — impressive when it works, opaque when it doesn't. This is infrastructure you stand up, not a one-click extension.

An AGPL framework (5.9k stars) of 12 pipelines, 52 tools, and 500-plus agent skills that teach a coding assistant professional video production — generation, image, TTS, music, post. It produces real edited cuts from stock or AI assets, not animated slideshows, with quality gates and budget controls along the way. Reach for it if you already drive Claude Code or Codex and want it to cut an explainer or talking-head without learning a separate video app. It fills the gap a standalone AI-video SaaS subscription would — except it runs inside the agent you already use. Tradeoff: this is the vertical extreme of the skills idea, and it shows — Python 3.10+, FFmpeg, Node, and paid API keys for full-quality output. Last meaningful commit was about six weeks ago, so treat it as a strong reference build to learn from, not a maintained dependency.

The Rust inference engine `mistral.rs` (MIT, 7.3k stars, v0.8.12) added a `/v1/skills` endpoint — OpenAI-compatible, so the same SKILL.md skills load against a model running on your own hardware instead of a vendor API. Reach for it when you want the skills workflow but with local models: air-gapped, cost-capped, or privacy-bound work where the cloud isn't an option. It deletes the assumption that skills are a cloud-vendor feature — the standard is portable down to the runtime now. Tradeoff: this landed in a point release straight from the maintainer and adoption is early, so the endpoint is real but you're on the frontier, not the paved road. Verify behavior against your own skills before wiring it into anything that matters.

The day's contrarian pick. `gortex` (Apache-2.0) is a code-intelligence engine that ships as a single binary with a CLI, MCP server, and HTTP API. It indexes your repo into a graph and hands an agent only the symbols it asks for, claiming up to 50x fewer tokens than dumping files into context. It covers 257 languages via tree-sitter, with compiler-grade resolvers for Python, TypeScript, Go, Java, C# and the rest, and runs 100% local — `curl -fsSL https://get.gortex.dev | sh`. Use it when your agent burns its window re-reading files to answer "where is this defined" or "who calls this." It replaces the habit of pasting whole files — or letting the agent grep blindly — into the prompt. Tradeoff: the 50x figure is the vendor's best case on large repos; on a small codebase the indexing overhead may not earn its keep. It's the inverse of today's other picks — the win here is subtraction, not capability.