education-agent-skills: a pedagogical knowledge layer shipped as plain folders

One folder per skill, with no build step

The unit of distribution here is a folder. Copy skill folders from the skills directory into your agent's skills directory, and each skill is a folder containing a SKILL.md file with name and description frontmatter, with no dependencies and no build step. That single sentence is the whole architecture, and it is why the project works in Claude, Codex, Hermes and anything else that supports the Agent Skills standard without any of them needing an integration. The badge in the README names the standard version, Agent Skills 1.0, and a second badge counts 165 skills. The library spans 20 domains, and the split is deliberate. Domains 1 to 19 are teacher and designer facing, covering pedagogy, learning science, curriculum and assessment, while domain 20 is described as the first student facing domain, holding live AI interaction patterns that shape how an AI responds to learners during study sessions. A manifest in the repository root, registry.json, and an llms.txt file sit alongside the skills directory, which suggests the collection is meant to be indexed by tools as much as read by people.

The README points educators and builders at different doors

Two audiences are addressed separately, and the split shows in how the documentation is organised. For educators it says no setup required, with the plugin, a local skill install and manual copy and paste all offered as ways to start teaching. For developers and AI builders it points at the architecture instead, listing YAML schemas, typed inputs and outputs, chaining metadata and a live MCP server. The audience list underneath is more specific than a generic education audience would be. Classroom teachers who want evidence based lesson and assessment design without hours of research sit next to university lecturers who received little or no teacher training, curriculum designers and heads of learning, school leaders in international, Montessori, project based and nature based school settings, innovators building new school models, EdTech developers who need a programmatically accessible education knowledge layer, and education researchers interested in how evidence translates into AI mediated practice. The manual route is the floor of the design: opening any skill file, copying the prompt block and pasting it into any AI takes no tooling at all.

Installing into Claude and Codex

Each agent gets its own invocation, and the commands are short. In Claude Code the install comes from the repository URL:

claude plugin install https://github.com/GarethManning/education-agent-skills

For Claude there is also a plugin path through Customize and the add plugin option, where you paste the same repository URL, plus a hosted MCP connector for Claude.ai and Claude Desktop as an option only for workflows that need a remote connector. Codex takes a different route, because Codex does not need the hosted MCP server, so the local setup is clone, change directory, and add the directory as a marketplace:

git clone https://github.com/GarethManning/education-agent-skills.git
cd education-agent-skills
codex plugin marketplace add "$PWD"

A Codex plugin manifest at .codex-plugin/plugin.json points at the skills directory, a local marketplace helper lives at .agents/plugins/marketplace.json, and Codex needs a restart after the plugin is enabled. For one or two individual skills, the copy is explicit, and the README gives a concrete path rather than a placeholder:

cp -r skills/memory-learning-science/spaced-practice-scheduler ~/.codex/skills/

The hosted MCP server is now token gated

The most consequential change in this repository is announced in a highlighted note at the top of the README. Hosted MCP access now requires an auth token, and anonymous access is blocked so that, in the project's words, the service stays sustainable. The library itself remains free and open source, and local, plugin and manual use is still called the recommended free path. The hosted server is framed as a convenience endpoint for remote clients, not a requirement for using the library, and there is a request access page for people who need it. The endpoint is a plain URL:

https://mcp-server-sigma-sooty.vercel.app/mcp

The two tools the hosted route adds are find_skills and suggest_skills, which give intelligent skill discovery and recommendation instead of local or offline installs. The README is careful to say that the MCP route and the Hermes tap serve different adopter types, and that a separate Hermes plugin is not currently planned. Reading that last clause is a useful signal about where the project's own energy is going.

Hermes gets a tap, and a warning not to install everything

Hermes users are told to keep the repository as the canonical source and then install only the skills they actually need locally, which is the same philosophy stated for the other agents. Two commands do the work, one adding the tap and one installing a named skill:

hermes skills tap add GarethManning/education-agent-skills
hermes skills install \
  GarethManning/education-agent-skills/skills/original-frameworks/learning-target-authoring-guide \
  --category education --yes

The guidance that follows is the most opinionated line in the documentation. The recommended starting point is to install selected skills or a small starter pack rather than all 165, and the reason given is that this keeps your local Hermes index useful instead of noisy. That is a useful piece of advice for every agent, since the same index problem exists wherever a tool discovers skills by scanning a directory. The full Hermes guide lives in the documentation directory at docs/HERMES.md, and a fuller Codex guide at docs/CODEX.md, with hosted access described separately in docs/HOSTED_MCP_ACCESS.md.

What else is in the repository

The tree shows a project with more machinery than a folder of prompts. There are separate plugin manifests for Claude, at .claude-plugin, and for Codex, at .codex-plugin, alongside the .agents directory, and the skills directory itself is organised by domain, since the copy command points at a real path inside the skills tree, skills/memory-learning-science/spaced-practice-scheduler. An mcp-server directory holds the hosted server implementation, and a scripts directory and a tests directory sit next to it. The root carries a package.json whose only script is a Playwright test run, with yaml and the Playwright test package as the development dependencies, and a playwright.config.ts at the top level, so the automated checks are browser driven. There are also documents that are unusual for a prompt collection: AUDIT.md, STATE.md, CHANGELOG.md, AGENTS.md and CLAUDE.md, alongside CONTRIBUTING.md and a directory of assets. The presence of STATE.md in particular suggests the project tracks its own progress as a file rather than only in commit history.

Licence, the argument for a research base, and the release channel

The library is published under CC BY-SA 4.0, which the badge states and the package manifest repeats, although the repository licence field is recorded as a custom licence that GitHub cannot classify, so the LICENSE file is the authoritative text. The case the project makes for itself concerns the base that AI education tools are built on. The README argues that whether AI improves learning outcomes or simply scales mediocre practice depends almost entirely on what it is built on, and it describes most AI education tools as built on convention, habit and assumption rather than on what the research says works, naming learning styles, rigid lesson structures and wellbeing programmes disconnected from learning theory as the habits in question. The library is positioned as an alternative foundation, anchored in named research and honest about limitations, aimed at people building the next generation of schools rather than optimising existing ones. On maintenance, the last push was on 2026-08-28, the repository is not archived, and there are no GitHub releases, so the CHANGELOG.md file is where the history lives and the branch tip is what you install.

Editorial conclusion

education-agent-skills fits a teacher who wants a research grounded lesson or assessment prompt today and can paste it into any assistant, and it fits an EdTech builder who needs a structured education knowledge layer that is inspectable rather than hidden behind an API. It is the wrong choice if your workflow depends on a hosted connector, since anonymous access to the MCP server is blocked and a token must be requested, and it is a poor fit for anyone who wants a single plug-in behaviour, since the recommended install is a subset rather than the whole library. Before you adopt it, read the README description of domain 20 and decide whether student facing content belongs in your product, and check the CHANGELOG.md, because the project publishes no GitHub releases to pin against.

Frequently asked questions

Official sources

1. GarethManning/education-agent-skills on GitHub
2. Issues
3. Project website
4. README