Skip to content

OKF directory & marketplace strategy

Decision-ready as of 2026-07-29. OKF was announced 2026-06-12; every number below is a snapshot of a seven-week-old ecosystem and will move fast. Claims are tagged [verified] (fetched from the cited source) or [inference] (our read). Where a source could not be reached from this environment, that is said explicitly.


Executive summary

Do not build a hosted bundle marketplace; ride GitHub as the registry and ship the index. Outcrop should publish an index-of-bundles that is itself an OKF bundle — curated in a public repo, consumed in-product through the existing canvas (a query card and a graph card over the index), and installable via a new outcrop install <owner/repo> command — so browsing the directory is using the product. The ecosystem already has two community web directories, three awesome-lists, and five-plus OKF MCP servers, but nobody validates OKF's trust metadata (sources, generated, verified, stale_after); Outcrop is the only tool positioned to make trust queryable (WHERE _trust_tier = 'human-reviewed' AND NOT _stale), and that is the differentiator to build the community around. Distribution rides existing rails — the okf-bundle GitHub topic (only 3 repos today; the namespace is unclaimed and we should write its convention), MCP tools (install_bundle, search_bundles) into an 80k-server MCP distribution wave, and canvases/skills/themes that already persist as records and therefore ship as bundles for free. Revisit this posture only if Google ships an official registry (integrate it day one) or a community directory reaches thousands of bundles with a stable API (consume it, don't compete with it).


Part 1 — The OKF ecosystem as of July 2026

1.1 The spec and its owner

  • Google Cloud announced OKF v0.1 on 2026-06-12 with reference implementations (a BigQuery enrichment agent producer, a static HTML visualizer consumer), three sample bundles (GA4, Stack Overflow, Bitcoin) living in-repo at okf/bundles/, and Knowledge Catalog ingestion. The announcement names no partners and no registry; governance is "file issues, send PRs" on GitHub. [verified] — announcement, okf/bundles
  • GoogleCloudPlatform/knowledge-catalog: ~8,000 stars, 664 forks, 88 open issues, 66 open PRs as of 2026-07-29. The README carries the disclaimer "not an official Google product." [verified] That is real momentum for seven weeks, and real governance ambiguity: the spec's home is a samples repo, not a standards org.
  • The spec moved fast already: v0.2 (which we vendor at docs/okf/SPEC.md, fetched 2026-07-29 per docs/okf/UPSTREAM) supersedes v0.1 with two breaking renames (timestamp → generated.at; body # Citations → sources) plus the trust/provenance/attestation families. Most third-party tooling surveyed below still targets v0.1. [verified against the vendored spec §13; v0.1 targeting observed in tool READMEs]
  • okf.md renders the spec and hosts an ecosystem tools page with producer/consumer categories. The site itself was unreachable from this environment (proxy 403), so its maintainer and listing process are unverified; its existence and structure are confirmed via search index.

1.2 Directories that already exist

Directory What we could verify What we could not
BundleDex Community registry claiming 240+ bundles; submission = "push to GitHub"; per-bundle pages; a "Using OKF bundles with Claude Code" guide; an okf get <name> fetch flow. [verified via search snippets] Site blocked from this environment (403). Owner, curation bar, traffic, and whether okf get is their own CLI — all unverified.
okfbundle.com Public directory plus a hosted MCP endpoint (okfbundle.com/mcp/) — point Claude Desktop/Cursor at one URL to search and read bundles. [verified via search snippets] Same: owner, count, curation unverified (403).
awesome-lists Three independent awesome-okf lists: linyiru (published as a conformant OKF bundle — the dogfooding move already exists in the wild), McClawdDigital, yzfly (Chinese; Feishu/Obsidian/Notion/GitHub → OKF converters). [verified] Relative reach.

Nobody official: Google hosts sample bundles in the spec repo and explicitly did not ship a registry. The directory layer is entirely community, entirely young, and nothing has won. [verified]

1.3 Tooling adoption signals

  • GitHub topic okf: 178 repos. Top by stars: lineage-skill (378★, distills media into source-backed agent skills), okf-skills (180★, Claude Code plugin to author/validate/visualize bundles), echoes-vault-opencode (153★, OpenCode persistent memory), claude-mega-brain (122★, OKF context injection for Claude Code), okf-gem (112★, CLI: validate | lint | search | index | server | render, per-user @slug registry in a plain JSON file, ships as a Claude Code plugin). [verified]
  • Topic okf-bundle: 3 repos. The bundle-tagging namespace is effectively unclaimed. [verified]
  • MCP is the hottest OKF distribution surface: at least five independent OKF MCP servers in seven weeks — hdean-ssp/okf-mcp (BM25+vector search over bundles), OKF Vault, OKF Knowledge Agent (kb_query/kb_add/kb_update tools), mfdaves/okf-mcp, concept-router, plus okfbundle.com's hosted endpoint. [verified]
  • Vendor support: secondary roundups (e.g. witscode) claim Ab Initio, Anomalo, Atlan, and Collibra announced OKF support; Google's own announcement names none. Treat as plausible but unverified. An Obsidian interop commit exists in the wild (obsidian-wiki OKF import/export); Obsidian needs no plugin to read a bundle — it is just a vault. [verified]
  • HN: "Google proposes Open Knowledge Format based on Markdown". Thread unreachable from this environment; secondary coverage (iloveseo, webreactiva) reports the dominant objections were (a) no typed relationships between concepts and (b) "weird text format + git" is hostile to non-dev knowledge workers. Both objections are things Outcrop directly answers (graph table functions; a DBMS/canvas surface over the files). [objections verified via secondary sources; thread stats unverified]

Read on the field [inference]: the ecosystem is accreting through agent clients (Claude Code plugins, MCP servers, OpenCode memory), not through web directories. The two directories exist but neither has structural advantage; the awesome-lists show curation demand; the trust/provenance half of the spec — v0.2's entire reason to exist — is validated by no tool we found.


Part 2 — How ecosystems actually accrete users

Full survey in Appendix A. The mechanics that repeat:

  1. The client is the directory that wins. Obsidian's plugin list is a git repo plus an in-app browser; Raycast's store is in-client; Hugging Face won because from_pretrained("name") was wired into the library everyone already used; Docker Hub won by being the client's default registry. Standalone web directories commoditize instantly — see the Claude-skills scrape-fest, where five sites each claim to be the largest (23k, 42k, 96k, 274k, 2.4M "skills") and none is trusted.
  2. Publish friction must be ~zero, and the publisher keeps hosting. Homebrew taps, Go modules, Obsidian plugins, BundleDex's own "push to GitHub to submit": the registry that asks for an upload loses to the registry that asks for a URL. Hosting bytes is a cost center that later forces rate limits (Docker Hub's arc).
  3. First-party blessing sets the namespace even when it arrives late. The official MCP registry appeared after Smithery/Glama/PulseMCP had tens of thousands of listings (~81.8k deduplicated across five sources as of 2026-07-28 per mcptoplist) — and became the canonical metadata source the aggregators now consume.
  4. When counts inflate, trust becomes the scarce good. The MCP and skills ecosystems are both now differentiating on curation, signing, and provenance — after the fact. OKF v0.2 bakes the trust fields into the format itself; the ecosystem just hasn't operationalized them yet.

Part 3 — Strategy

3.1 Posture: ride GitHub as the registry, ship the index

Recommendation: community-first, first-party index. Outcrop does not host bundles, does not run a marketplace website, and does not compete with BundleDex on breadth. Outcrop publishes and curates outcrop-index — an index of known-good OKF bundles that is itself a conformant OKF bundle in a public repo:

  • Each directory entry is a concept: type: Bundle, resource: https://github.com/<owner>/<repo>, description, tags, sources (where we found it), and — the point — verified entries recording that Outcrop's validator ran clean against it, at what commit, by which actor (process:outcrop-index-ci or human:<curator>).
  • Curation is PR-based (the Obsidian model). CI on the index repo runs outcrop validate against each listed bundle before merge.
  • Because the index is an OKF bundle, it is: consumable by every other OKF tool (including the directories we're "competing" with — they can ingest us), diffable, forkable if we ever misbehave, and — decisively — queryable in Outcrop itself. The in-product bundle browser is not new UI; it is a canvas with a query card and a graph card pointed at the index bundle (per docs/CANVAS.md: cards are queries, the canvas is a record). Browsing the directory is a live demo of the product. Dogfooding is the moat.

What Outcrop builds:

  1. outcrop install <owner/repo>[/path][@rev] — fetch a bundle from GitHub (or any git URL/tarball) into a folder of the store, through the normal transaction path: staged, validated per folder policy, committed with History. Install provenance is recorded in OKF's own vocabulary — a sources entry on the imported records' folder metadata pointing at origin repo + commit. Uninstall = Restore.
  2. outcrop publish — validate + lint + generate index.md + tag the repo okf-bundle + open a PR against outcrop-index. A helper, not a gate: publishing is pushing to GitHub.
  3. MCP tools search_bundles and install_bundle on the existing outcrop-server MCP surface — so any agent in any MCP client can pull knowledge into a store mid-session. This rides the largest distribution wave in the agent ecosystem at zero marginal cost.
  4. Trust made queryable — computed columns in outcrop-query: _trust_tier (unverified | machine-confirmed | human-reviewed, derived per spec §5.3), _stale (today >= stale_after, §5.5), _provenance_depth (link-recursion through sources, §5.1). Plus outcrop verify <folder>: a trust report over any bundle. No surveyed tool does this. It converts OKF's paper promise ("agent-maintained corpora you can trust") into a SQL WHERE clause, and it is exactly Outcrop's constraint philosophy: a low-trust record is a valid file and a flagged record, never an error.

What Outcrop does NOT build:

  • A hosted marketplace, bundle storage, or CDN. (Cost center; loses to GitHub.)
  • Ratings, reviews, comments, or a social layer. (The skills directories show where that goes without an install base.)
  • Payments/monetization. (Notion-template economics require consumer scale we don't have; revisit only with evidence.)
  • A Outcrop-specific packaging format, lockfile, or manifest beyond what OKF specifies. (BRIEF #9's spirit: no bespoke conventions; the bundle is the package.)
  • A fork or extension of the spec. We propose upstream via PR or we conform.

Revisit triggers:

  • Google ships an official OKF registry → integrate it day one as the primary upstream of outcrop-index; our value shifts entirely to validation + client.
  • Any community directory passes ~1,000 bundles with a stable public API → consume it as an index upstream; do not out-breadth it.
  • Users demand private/team bundle sharing → that is a hosted product (private index + auth), a revenue decision, not a community one.
  • The okf-bundle topic convention fails to take (< 50 repos by Q4 2026) → fall back to scanning okf topic + awesome-lists for index candidates.

3.2 Ubiquitous integration plan (sequenced)

  1. Claim the tagging convention. Write okf-bundle topic guidance (tag = "this repo/subdir is an installable OKF bundle; root has index.md with okf_version"), publish it in the index repo, tag our own sample corpus, and PR the three awesome-okf lists to adopt it. Three repos use the topic today; whoever writes the convention owns the namespace. [the 3-repo count is verified; the opportunity is inference]
  2. Be present everywhere bundles are listed. Submit Outcrop to okf.md/tools, all three awesome-okf lists, BundleDex, and okfbundle.com; list outcrop-server on the official MCP registry, Glama, PulseMCP, and Smithery. Cheap, compounding.
  3. MCP first, GUI second. search_bundles/install_bundle tools land before the canvas browser; agents are the growth surface with momentum.
  4. Everything Outcrop makes is a bundle. Canvases, skills, and themes already persist as records (type: canvas|skill|theme, per docs/CANVAS.md) — so a dashboard, an agent skill, or a theme is shareable through the same install path as data. "Share this canvas" = publish a bundle. This is the community flywheel no data-catalog vendor can copy: their OKF export is documentation; ours is working software (a dashboard that renders on install).
  5. Obsidian is interop, not integration. A bundle is a valid vault; the demo loop (edit in Obsidian → watcher reindexes) already exists in DEMO.md. Ship a one-page guide, not code.
  6. Upstream presence. Watch the 66 open PRs on knowledge-catalog; file conformance issues we hit; propose nothing until we've shipped validators that give us standing.

3.3 Risks

Risk Read Mitigation
Google controls the spec — and hosts it in a repo marked "not an official Google product." Governance could formalize (good), stall (fine), or swerve toward Knowledge Catalog lock-in (bad). Medium probability, high impact Already partly handled: okf-spec-watch workflow + vendored spec + conformance tests. Add: never depend on fields beyond the spec; keep our extensions in producer-defined keys (spec §4.1 permits).
Spec churn. v0.1→v0.2 broke two fields in five weeks. Most ecosystem tooling still emits v0.1. Certain, low impact each Consume v0.1 via the spec's own fallbacks (§13: timestamp, # Citations); emit v0.2. Test both in the corpus.
Directory cold-start. An empty index convinces nobody. Certain, self-inflicted if we build a marketplace This is why the posture is ride-the-rails: the index launches pre-populated from the okf topic scan, Google's three sample bundles, and our flagship bundles — value on day one without any publisher adopting us.
Trust/provenance of third-party bundles. An installed bundle is agent-facing text: a malicious Playbook concept is a prompt-injection payload the moment an agent card reads it. Attested Computation executor/attester resources are code references. High impact, underappreciated ecosystem-wide Install-time trust report (tiers, staleness, unverifiable sources) before commit; never auto-execute executor/attester resources; folder policy warn|block applies to imports; provenance stamped on every install so WHERE clauses can quarantine by origin. Spec-conformant: we flag, we never reject (§11).
Directory capture — BundleDex (or successor) becomes the de-facto gate. Low-medium Our index consumes multiple upstreams and is itself an open, forkable bundle; we integrate winners rather than fight them.

Part 4 — 30/60/90-day actions

Sizing: S ≤ 2 days, M ≤ 2 weeks, L > 2 weeks. Owner: eng / product / community.

Days 0–30 — claim the rails

# Action Size Owner
1 Write + publish the okf-bundle topic convention; tag our sample corpus repo S community
2 outcrop install <git-ref> MVP (clone/subdir → folder via transaction path, provenance stamp, Restore = uninstall) M eng
3 MCP tools search_bundles (over local index copy) + install_bundle M eng
4 Computed columns _trust_tier, _stale in outcrop-query; outcrop verify report M eng
5 Submissions: okf.md/tools, 3 awesome-okf lists, BundleDex, okfbundle.com; MCP registries S community

Days 31–60 — ship the index

# Action Size Owner
6 outcrop-index repo: index-as-OKF-bundle; seeded from okf-topic scan + Google samples; CI validates every listed bundle; PR-based curation M product + community
7 outcrop publish (validate + lint + index.md + topic tag + index PR helper) M eng
8 Bundle-browser canvas: query card + graph card over the installed index bundle M product
9 Three flagship bundles that only Outcrop makes shine (data + canvas + theme + skill in one bundle) M product
10 v0.1 ingestion fallbacks (timestamp, # Citations) with tests S eng

Days 61–90 — light the flywheel

# Action Size Owner
11 "Share as bundle" flow for canvases/skills/themes (select records → publish) L product + eng
12 Monthly community bundle showcase + contribution guide on the index repo M community
13 Obsidian interop guide (bundle = vault; edit-and-reindex loop) S product
14 Install-time trust report UX in CLI + canvas (tiers, staleness, origin) M eng
15 Checkpoint: measure okf-bundle topic count, index PRs from non-team authors, directory traffic; evaluate revisit triggers (§3.1) S product

Appendix A — Surveyed directories & ecosystems

Name Unit of distribution Host Curation Adoption signal (as of survey) Lesson for Outcrop
knowledge-catalog (OKF home) OKF bundle (dir of .md) GitHub Google maintainers; PRs ~8k★, 664 forks, 66 open PRs in 7 weeks [verified] The spec is hot; the registry seat is empty.
BundleDex OKF bundle Publisher's GitHub; site indexes Submit-by-push; bar unknown 240+ bundles claimed [unverified — site unreachable] Third parties already ride GitHub; breadth is not defensible.
okfbundle.com OKF bundle Site + hosted MCP endpoint Unknown Hosted MCP browse/search [partially verified] MCP endpoint as directory surface is the interesting move — copy it.
Official MCP registry + aggregators MCP server (metadata) Publisher hosts; registry = metadata Namespace verification; aggregators scrape ~81.8k servers deduped across 5 sources [verified via mcptoplist] Late first-party blessing still wins the namespace; aggregators then consume it.
Claude/agent skills directories (skillsmp, agentskill.sh, claudemarketplaces, …) SKILL.md folder GitHub; sites scrape None visible Five sites claiming 23k–2.4M skills [verified claims, dubious counts] The anti-pattern: uncurated web directories commoditize to zero trust. Skills folder ≈ OKF bundle cousin — same fate awaits OKF directories.
Obsidian community plugins Plugin (GitHub repo) GitHub repo of listings + in-app browser PR review by Obsidian team ~6.1k plugins, 650 themes, 120M+ downloads [verified] Registry = a git repo + the client as storefront. Directly copyable.
Hugging Face Hub Model/dataset repo HF (git-LFS) Open upload; org verification; cards ~2M models [approximate, from general knowledge — unverified this session] The directory wired into the default tool (from_pretrained) wins the category.
Homebrew Formula/tap (git repo) GitHub Core = PR review; taps = anyone De-facto macOS package layer [general knowledge] Decentralized publish (tap <user/repo>) + curated core: exactly the index-plus-GitHub split.
npm Package (tarball) npm registry (hosted) None at publish; post-hoc security Millions of packages [general knowledge] Hosting everything = paying for everything + supply-chain liability. Avoid.
Docker Hub Image Hub (hosted) Official images curated Default-registry status won it; rate limits later pushed users to ghcr [general knowledge] Default-in-client wins; hosting costs eventually tax the ecosystem.
Notion template gallery Template Notion Marketplace review; paid tier Creator economy at consumer scale [general knowledge] Monetization needs consumer scale; premature for OKF.
Raycast store Extension Raycast (hosted) First-party review In-client, one-click [general knowledge] Small curated + in-client beats large uncurated for tool ecosystems.
awesome-lists (linyiru et al.) Link GitHub Maintainer taste 3 OKF lists in 7 weeks; one is an OKF bundle [verified] Curation demand exists; the index-as-bundle idea is already validated in miniature.

Appendix B — Source notes

Unreachable from this build environment (proxy 403, so claims about them rely on search-index snippets): bundledex.net, okfbundle.com, okf.md, the HN thread (id 48517735). Everything cited from github.com, cloud.google.com, and the aggregator/blog links above was fetched directly or returned in search results during this session (2026-07-29).