roadmap
Where we're going.
cache.directory is a long bet on structured, machine-readable information about the agent ecosystem. Here's what's shipped, what's in progress, and what comes next.
Phase 1 The Registry shipped
A swarm-produced, schema-validated, statically-served directory. The foundation everything else builds on.
✓ Six curated shelves: agent skills, MCP servers, AI starters, system prompts, editor rules, local LLM tools
✓ Zod-validated content schema — invalid entries go to quarantine, never the build
✓ Static Astro site on Cloudflare Pages — zero server, global edge, instant builds
✓ Machine-readable API:
/api/v1/:shelf.json, /llms.txt, /sitemap-index.xml✓ Compatibility matrix per entry (claude-code / cursor / cline / aider)
✓ Safety sandbox: Docker-isolated behavioral verification with verdict badges
✓ Swarm pipeline: AI operator agents running on cron, sourcing from GitHub / npm / PyPI
✓ Terminal-brutalist design system — dark-mode-first, mono typography, amber/green/rose palette
Phase 2 The CLI in progress
npx cache add anthropic-bash — a zero-dependency package manager for the agent layer. Resolves entries from the registry, places artifacts in the right place, writes a lockfile.
✓
GET /api/v1/resolve/:slug.json — registry resolve endpoint✓
GET /api/v1/raw/:shelf/:slug — raw artifact download endpoint✓
GET /api/v1/search.json — full-corpus search index✓
@cache/cli — zero-dependency Node 18+ CLI: add, list, info, search✓ Shelf-aware placement: skills→SKILL.md, mcp→.mcp.json (merge), rules→.cursorrules
✓
cache.lock TOML lockfile with sha256 verification✓ Error taxonomy: CONFLICT, MCP_KEY_COLLISION, NOT_FOUND, NETWORK_ERROR
○ Publish
@cache/cli to npm — making npx cache add work from anywhere○
cache upgrade — check installed entries against latest registry, diff and update○ Multi-skill projects — named skill files in a
./skills/ directory instead of project root○ Offline mode + local cache —
~/.cache/registry/ TTL cache for air-gapped use Phase 3 Scale planned
Expand the corpus to 4,000+ entries across 13 shelves. Add community contribution paths. Open the safety sandbox to third-party tools.
○ 7 new shelves: workflows, evals, benchmarks, fine-tunes, datasets, agent frameworks, vector stores
○ 4,000+ total entries — Wave 2 swarm campaign targeting long-tail tools
○ Community contribution path — curated PRs with swarm-assisted review
○ Entry versioning — semver on content,
cache upgrade diffs○ Organization-scoped private shelves —
@yourorg/cache○ GitHub Actions integration —
cache restore in CI from cache.lock○ Safety sandbox open to community — any tool author can submit for verification
Phase 4 The Standard vision
cache.directory becomes the canonical reference for what's safe to run in an agent environment. The CLI becomes infrastructure.
○ Editor plugins — VS Code / Cursor / Zed: browse and install from the registry inline
○ Agent-native install — MCP tools that can call the registry to discover and install other tools
○ Vendor-neutral safety standard — a published spec for agent tool behavioral verification
○ Automated freshness — swarm re-validates entries on upstream change events
○ Registry federation — other organizations can run compatible registry nodes
Not on the roadmap
Keeping the list of things we're explicitly not building is as important as the roadmap itself.
✗ Paid listings or sponsored placements — ever
✗ User accounts required to browse the registry
✗ Tracking pixels or analytics SDKs on the public site
✗ Closed/authenticated endpoints on the core registry API
✗ Binary formats or proprietary schemas — everything stays Markdown + JSON