Expert guidance for AI coding agents, delivered as canonical Markdown packs and generated Claude-compatible skills.
Agent Hub now has two related artifact types:
- canonical versioned packs in
agents/<tool>/<version>.md - generated distributions built from those packs, including Claude-compatible skills in
distributions/claude/<tool>/<version>/
The canonical pack is the source of truth. Generated distributions are delivery formats built from that source.
This gives Agent Hub two practical delivery paths:
- runtime retrieval over MCP
- file-based installation through generated Claude-compatible skills
The canonical pack is the authoritative authored artifact:
agents/<tool>/<version>.md
It is:
- versioned
- validated
- reviewed directly in GitHub pull requests
- the thing contributors should edit
The generated Claude-compatible skill is a delivery artifact:
distributions/claude/<tool>/<version>/
SKILL.md
references/
manifest.json
It is:
- generated from the canonical pack
- designed for progressive disclosure in Claude-compatible environments
- a distribution target, not the primary authoring surface
Default rule:
- edit the canonical pack
- regenerate the Claude-compatible skill
- do not hand-edit generated skill files unless the repo explicitly says otherwise
Use Agent Hub MCP when the agent should:
- discover available packs
- inspect versions
- fetch the right pack only when a task needs it
- keep prompts smaller and more repeatable
Start here:
Use the generated Claude-compatible skill when the environment expects:
- a local
SKILL.mdbundle - file-based installation
- progressive disclosure through a short skill entrypoint plus supporting files
Start here:
Agent Hub is built in public through GitHub pull requests.
Good starting points:
If you are contributing packs, remember:
- canonical pack first
- generated distribution second
- regenerate rather than hand-edit generated skill bundles
- a pack-generation task should normally produce both the canonical pack and the matching Claude-compatible skill
Validate a canonical pack:
npm run validate:agent-pack -- agents/<tool>/0.4.0.mdGenerate a Claude-compatible skill bundle:
npm run generate:claude-skill -- agents/<tool>/0.4.0.mdFinalize a pack the default way:
npm run finalize:agent-pack -- agents/<tool>/0.4.0.mdGenerate the current pilot set:
npm run generate:claude-skill:pilotCheck generated Claude-compatible bundles for drift:
npm run check:claude-skillInstall a generated Claude-compatible skill into .claude/skills/ in the
current project:
npm run install:claude-skill -- agent-hub --projectInstall a generated Claude-compatible skill into ~/.claude/skills/:
npm run install:claude-skill -- react --globalCheck the MCP distribution tools locally:
npm run check:mcp-distributionsCheck Claude Code local skill install + MCP setup preflight:
npm run check:claude-code-local-setupRun the logged-in Claude Code runtime validation probes:
npm run check:claude-code-skill-runtimePost the resulting runtime validation report straight to issue #13:
npm run check:claude-code-skill-runtime -- --issue 13Build the full site and synced static artifacts:
npm run buildagents/— canonical versioned packsdistributions/claude/— generated Claude-compatible skill bundlesparse/— intermediate documentation packsprompts/— generation prompts and runbooksspec/— normative format specsscripts/— validators, generators, and sync toolingtutorials/— user and contributor workflow guideswebsite/— Docusaurus sitenetlify/functions/mcp.js— deployed MCP server