AI Agent Registries & Directories: the 4 That Matter (and the Junk Tier)

In short: Only four agent-registry placements show measured impact as of 2026-06-11. They are the Official MCP Registry (which feeds GitHub's registry and renders natively in VS Code), skills.sh (~1.38M weekly CLI downloads), Anthropic's Connectors Directory and plugin marketplace, and Context7 (~1.14M weekly npm downloads). About 99% of real distribution bypasses standalone directories. List in those four. Skip the rest.

Works in: Find & Research. Registries get you indexed (so you show up in search) and backed up (a second party naming you).

Do this now

• Publish to the Official MCP Registry via mcp-publisher (free, ~1 hour). Confirm your entry renders in VS Code's Extensions view under @mcp.
• Make your skills repo public with SKILL.md files. skills.sh lists it automatically, with no submission. Then embed the badge in your README.
• File your Anthropic Connectors Directory submission today (review takes days–weeks) and prepare a .claude-plugin/marketplace.json listing.
• Submit your docs site (not your repo) to Context7 at context7.com/add-library. Claim the entry and configure parsing.
• Rewrite your Context7 description to contain the exact category words developers type. First check who wins those queries today.
• Spend ≤2 hours on Tier 2: PulseMCP, Smithery, and a Cursor "Add to Cursor" deeplink on your own docs.
• Skip mcp.so, Glama, mcpmarket, awesome-lists, mcp-get, and ClawHub. Write down that you skipped them, so nobody "fixes" it later.
• Stand up the weekly position tracker (below) before calling this play done.
• (Play 12, frontier) If you ship an MCP server or API, draft a /.well-known/ai-catalog.json for Google's new ARD spec. It's a ~1-hour first-mover option; 0 of 39 top sites publish one yet.

📥 Free resource: Registry Submission Runbook worksheet

Who needs this play?

Anyone who ships an MCP server (Play 3), a skills repo (Play 4), or docs that developers integrate against. Registries are where agents and their host clients resolve "what exists" for a task. But only a handful feed real traffic, so most of this play is about where not to spend time.

Scope note: this play covers the structured indexes agents pull from on purpose (Context7, the MCP registry). For the open-web search an agent runs itself (appearing in its results, surviving the fetch, and getting past its verification), see Play 1. They are complementary discovery surfaces.

Prerequisites: a public GitHub org, a docs site, and DNS control for namespace verification. Budget 1–2 days for Tier 1, then about an hour a week of tracking.

Why do ~99% of installs bypass standalone directories?

One data point makes the whole landscape clear. As of 2026-06-11, Context7's MCP client (@upstash/context7-mcp) does 1,136,447 npm downloads per week. On Smithery, the best-known standalone MCP directory, the same product shows just 6.8K "uses." About 99.4% of its real distribution bypasses standalone directories entirely.

Distribution flows through npm, native client surfaces (VS Code's Extensions view, Claude Code's /plugin), and vendors' own docs funnels. Standalone directories mostly mirror demand that already exists. They rarely create it.

So we tiered every registry strictly by evidence: native client integration > published measured usage > third-party traffic estimates > GitHub stars > self-reported server counts (the last is worthless alone). Four placements survive the cut.

Which AI agent registries actually matter?

How do you get listed in the four Tier-1 surfaces?

1. Official MCP Registry: the upstream pipe (~1 hour)

Not because users browse registry.modelcontextprotocol.io. They don't. Its documented consumers include Smithery, PulseMCP, Docker Hub, Anthropic, and GitHub. The GitHub MCP Registry sources from it, and VS Code/Copilot renders the GitHub MCP Registry natively in the Extensions view. One free listing propagates downstream into the only registry a major client queries by default.

1. Install the mcp-publisher CLI (github.com/modelcontextprotocol/registry).
2. Pick a namespace: io.github.<org>/<server> (verified via GitHub OIDC) or reverse-DNS com.<yourdomain>/<server> (DNS TXT or HTTP challenge).
3. Write server.json: name, description, package or remote endpoint, version.
4. Run mcp-publisher login github, then mcp-publisher publish.
5. Confirm the entry appears in the GitHub MCP Registry, then search @mcp <product> in VS Code's Extensions view.
6. Re-publish on every release so downstream consumers carry current metadata.

Enterprise note: organizations can repoint VS Code's gallery via the McpGalleryServiceUrl policy. But the default gallery is what most users see.

2. skills.sh: there is no submission step

Any public GitHub repo containing SKILL.md files becomes installable via npx skills add owner/repo, and the leaderboard auto-populates from install telemetry. Your work is repo quality (Play 4 covers it), plus embedding the badge (https://skills.sh/b/owner/repo) in your README.

Set expectations correctly: installs are driven by your own funnels (docs, README, launch posts), not by directory browsing. The leaderboard is the scoreboard, not the channel. Counts are telemetry-based and cache-noisy, so treat them as order-of-magnitude.

3. Anthropic surfaces: reviewed, and not redundant

Submit to the Connectors Directory (claude.com/connectors; ~439 reviewed connectors, queried in-product) and prepare a Claude Code plugin marketplace listing. claude-plugins-official ships pre-loaded in Claude Code's /plugin Discover view.

The trap: Claude Code's /mcp does not browse the Official MCP Registry. Being in the official registry buys you nothing on the Claude surface. These are separate, parallel placements. Budget days-to-weeks for review and meet the security and annotation requirements before submitting.

4. Context7: the one directory with a quality loop you control

Context7's install base is real, measured demand for docs retrieval. Unlike every other directory, you can tune your entry:

1. Submit the docs site, not the GitHub repo, at context7.com/add-library. Site entries outscore repo entries. Convex's site entry benches 91.6 vs 79.9 for its repo, an 11.7-point spread from source choice alone. Repos carry README and contributor noise.
2. Claim ownership and set the parse config: branch, include/exclude folders (exclude changelogs, blog, marketing pages).
3. Write custom benchmark questions: the questions a developer integrating your product would actually ask. An LLM jury re-scores your entry from your snippets on every parse.
4. Set a re-parse cadence tied to docs releases. Freshness correlates with benchmark score (Spearman −0.54 against log hours-stale; the freshest five entries we audited average 83.6 vs 72.3 for the stalest five).

How do you write a description that wins category queries?

Context7's lexical search matches names and descriptions. That makes the description field a ranking surface, and right now a largely uncontested one.

The receipt, verified 2026-06-11: on the query "payments", DodoPayments ranks #1 (11,647 snippets, benchmark 83). Stripe, with 265,284 snippets and benchmark 84.5, is absent from that query's top-10, because its entry is named simply "stripe." Smaller player, better description-to-query fit, category win.

Entire shelves sit empty, too. The query "accept crypto payment" returns a wasteland: the top snippet source is Crypto++, a C++ cryptography library, with no relevant product at all. Read that as category-vacancy evidence, not a payments story. Task-phrase lexical space is unclaimed across categories, and the first product to put the phrase in its description owns the shelf.

The runbook:

1. List the 5–10 category nouns and verb phrases for your space ("auth", "send email", "vector database", and the task phrases your users type).
2. Run each query on Context7 today and record who wins.
3. Rewrite your description to contain those exact terms, phrased naturally, not stuffed.
4. Re-check positions weekly (tracker below).

Which directories should you skip, and why is ClawHub a brand risk?

Tier 3: skip, or cap at five minutes for the backlink. mcp.so (~19.7K listed servers), Glama (~21–34K), and mcpmarket publish scrape-driven server counts with no client integration and no published usage. The awesome-mcp lists are star-counting. mcp-get is dormant. Hours in, nothing measurable out.

ClawHub: skip, and document why. OpenClaw's skill registry (~10.7–13.7K skills) carried 341 confirmed malicious skills distributing the AMOS stealer (The Hacker News, Feb 2026). Snyk's count ran to 1,467, and third-party estimates put 8–20% of the registry as malicious. Listing there creates brand-risk adjacency. Only consider it if OpenClaw users are specifically your market, and put that decision in writing.

Two non-directory notes while you're here. Codex has no registry at all. Configuration is config.toml only, and discovery is an open feature request (openai/codex#25750); your Codex placement is install snippets in your own docs. And Cursor's in-product MCP directory appears curated/BD-gated with no public self-serve process we could verify. Don't block on it; publish "Add to Cursor" deeplinks instead (cursor.com/docs/context/mcp/install-links), which work without any listing.

What's the full submission runbook?

How do you track registry positions weekly?

Automate five checks (cron or CI), append results to a dated CSV, and alert on any FAIL or two consecutive WARNs:

1. Official registry presence: GET registry.modelcontextprotocol.io/v0/servers?search=<name> returns your entry at the current version. FAIL if absent or stale by more than one release.
2. VS Code surface: the GitHub MCP Registry page for your server resolves. FAIL on 404.
3. skills.sh installability: npx skills add <owner>/<repo> exits 0 in a clean container; log the install count as a trend line (order-of-magnitude only, since counts are cache-inconsistent).
4. Context7 position: rank for each of your five target task queries, plus benchmark score and last-parse timestamp. PASS = top-3 on the primary term; WARN on a ≥2-position weekly drop; FAIL if the benchmark falls below 80 or staleness exceeds 14 days.
5. PulseMCP visitor estimate: secondary trend line only. npm weekly downloads of your own MCP package remain the honest ledger everything else calibrates against.

Quarterly, re-verify the tier table itself. This landscape reshuffled twice in the past year (ClawHub's rise and contamination, and the GitHub registry's VS Code integration), so the tiers have a shelf life. Full measurement design lives in Part 5.

What about ARD, the new open registry standard?

Everything above is the ad-hoc landscape as it exists today: a handful of vendor-owned directories with no shared protocol. In late May 2026, Google and a Linux Foundation working group proposed the standard that would replace the ad-hoc part: Agentic Resource Discovery (ARD). Instead of submitting to each directory, you would host one /.well-known/ai-catalog.json describing your MCP servers, agents, and APIs, and any ARD registry could crawl it.

If it gets adopted, ARD turns this play's tier table on its head: the open, self-hosted catalog becomes the upstream pipe, and directories become consumers of it (the way the Official MCP Registry already feeds GitHub and VS Code). If it doesn't, it joins the standards graveyard and this play stands as written.

Our read today: write the file (it's a cheap first-mover option), but don't restructure your distribution around it yet. As of 2026-06-18, our census found 0 of 39 major sites, including all eleven companies that authored the spec, publishing a discoverable catalog. The full field-by-field guide, the schema, and the census are in Play 12: Agentic Resource Discovery (ARD).

All figures verified 2026-06-11; methodology and updates in the Data Room.

Selection evidence (pilot-grade). In our pilot experiment E4 (single model, Haiku 4.5, n=2 per arm, 2026-06-11), telling an agent that one email provider ships MCP + llms.txt + skills flipped its provider selection 2/2; the control arm chose the other provider 2/2. Discovery surfaces are now a selection criterion, not just a convenience. Treat this as a pilot, not a benchmark.

Vendor-scale numbers on skills.sh (~95 official vendor orgs, including Stripe, Supabase, Clerk, and Auth0): Microsoft 4.5M installs aggregate, Convex 362.1K, Better Auth 115.5K, Resend 17.6K, DodoPayments 3.9K. Zero-gatekeeper listing, real measured numbers, but cache-inconsistent (shadcn's totals varied 63.9–133.9K between page loads), so order-of-magnitude only.

Counterexamples we audited. Crossmint ships a docs MCP, llms.txt, and agent-facing APIs, yet has no findable Context7 entry: full build cost, near-zero retrieval presence. Hono's repo-sourced Context7 entry has 48 snippets, benchmark 63.3, and is absent from the top-50 (a deliberate infrastructure positioning, but the retrieval invisibility is the measurable consequence).

Data-quality labels we use in every report: measured (npm downloads, skills.sh telemetry) > estimated (PulseMCP visitors) > vanity (Tier-3 server counts). "No usage evidence" for Tier 3 is absence of evidence, but after deliberate search, and the burden of proof is on the directory.

FAQ

Where should I list my MCP server? Four placements show measured impact: the Official MCP Registry (which propagates to GitHub's registry and VS Code), skills.sh for your skills repo, Anthropic's Connectors Directory plus plugin marketplace, and Context7 for your docs. PulseMCP, Smithery, and Cursor deeplinks are cheap Tier-2 additions. Everything else is skippable.

Is Smithery worth listing on? Yes, but for completeness, not distribution. It's free, takes about an hour, and offers hosted runtime/auth some users want. Measured usage is tiny, though (Context7 shows 6.8K uses there against 1.14M weekly npm downloads). List it. Don't report it as a channel.

How do I get into Context7's top results for my category? Submit your docs site (not your repo) at context7.com/add-library, then write your description to contain the exact category words developers type. DodoPayments ranks #1 for "payments" with a fraction of Stripe's corpus because its description matches the query. Keep the entry fresh; benchmark scores decay with staleness.

Is ClawHub safe to publish skills on? We recommend skipping it. Researchers confirmed 341 malicious skills distributing the AMOS stealer there (reported Feb 2026; Snyk counted 1,467; estimates run 8–20% of the registry), and the brand-adjacency risk outweighs the reach unless OpenClaw users are specifically your market.

Does Claude Code read the Official MCP Registry? No. Claude Code's /mcp does not browse the official registry; its /plugin Discover view ships with Anthropic's claude-plugins-official marketplace instead. Submit to Anthropic's surfaces separately. An official-registry listing does not cover the Claude surface.

---

Last verified 2026-06-11. We re-test the claims on this page quarterly; changes are logged in the Data Room.

Part of The Complete Playbook to Agentic Discovery.

Stay ahead of the agents. We re-test this playbook quarterly and publish what changed: new data, busted myths, ranking shifts. Get the update digest →

Want this done for you? Synscribe runs agentic-discovery programs for B2B SaaS and developer platforms. Talk to us →