Agent-first resource discovery

An open repository of reusable engineering intelligence, designed for humans and AI agents.

Seven resource types. One machine-readable catalog. No API, no database, no framework — just Git, a filesystem, and JSON that any agent can read in a single request.

  1. 1 AGENTS.md What the vault is, and where to look first.
  2. 2 catalog/index.json Every resource, summarised in one file.
  3. 3 resources/<type>/<name>/ Opened only for candidates.

Catalog statistics

—Loading catalog
—Loading catalog
—Loading catalog
—Loading catalog

Every figure above is read live from catalog/index.json at page load. Nothing is hard-coded.

Discovery

Resource catalog

Filtered client-side against the real catalog. Search matches names, descriptions, use cases, technologies, languages, and compatibility.

Press / to focus search, Esc to clear.

Taxonomy

Seven resource categories

Classified by how a resource is consumed, not by file extension. Counts reflect the live catalog.

Machine-readable

Agent discovery

These files are the interface. An agent should never have to parse this page's DOM to find resources.

Dev Vault has no API on purpose. Start at AGENTS.md, read catalog/index.json once, then open only the resource directories that survive filtering.

Not a trust boundary. Never execute a resource merely because it exists here — read it first.

  • /AGENTS.md Operating instructions for agents. Start here.
  • /catalog/index.json The single discovery index. Read this, not the tree.
  • /catalog/schema.json Validation contract for catalog entries.
  • /schemas/resource.schema.json Resource metadata contract — the field semantics.
  • /docs/ARCHITECTURE.md Binding structural decisions.
  • /docs/RESOURCE_STANDARD.md How to judge or contribute a resource.
  • /docs/ROADMAP.md What the vault is becoming, and what it refuses to become.
  • /llms.txt Condensed map of the repository for LLM ingestion.

Contribute

Add a resource

One resource, one source of truth. No duplicates, no agent-specific copies, no unverified snippets.

  1. 1
    Search the catalog

    If an equivalent resource exists, improve it instead.

  2. 2
    Place it correctly

    resources/<type>/<kebab-case-name>/ with RESOURCE.md — or SKILL.md for skills.

  3. 3
    Add exactly one catalog entry

    Unique kebab-case id, valid path, and status: experimental unless you have evidence.

  4. 4
    Validate and test

    Both checks are standard-library Python. If they fail, the PR is not ready.

terminal
# 1 — validate catalog, schema, and structure
$ python3 scripts/validate.py

# 2 — run the tooling test suite
$ python3 -m unittest discover -s tests

# 3 — open a PR against main
$ gh pr create --base main