Registry Submission Runbook

What this is: the working-checklist version of AI Agent Registries & Directories. Submit to the four placements that show measured impact, skip the rest in writing, and track positions weekly.

How to use it:

• Budget 1–2 days for the four Tier-1 checklists, then about an hour a week on the tracking grid. Prerequisites: a public GitHub org, a docs site, and DNS control for namespace verification.
• File the Anthropic submission first. Its days-to-weeks review runs while you finish the other three.
• Fill in the description worksheet before submitting to Context7, and record the skip-list decision so nobody "fixes" it later.

The tier table

Context: roughly 99% of real distribution bypasses standalone directories. Context7's client does 1,136,447 npm downloads/week against 6.8K "uses" on Smithery (as of 2026-06-11). Two no-registry surfaces to remember. Codex has no registry at all (config is config.toml only; discovery is open feature request openai/codex#25750), so your Codex placement is install snippets in your own docs. Cursor's in-product directory is curated/BD-gated, so don't block on it; publish "Add to Cursor" deeplinks instead.

Checklist 1 · Official MCP Registry (~1 hour, free)

Users don't browse registry.modelcontextprotocol.io, but its documented consumers include Smithery, PulseMCP, Docker Hub, Anthropic, and GitHub, and the GitHub MCP Registry it feeds renders natively in VS Code's Extensions view.

• Install the mcp-publisher CLI (github.com/modelcontextprotocol/registry).
• Pick a namespace: io.github.<org>/<server> or reverse-DNS com.<yourdomain>/<server>.
• Complete namespace verification: GitHub OIDC for io.github.*; DNS TXT record or HTTP challenge for reverse-DNS.
• Write server.json: name, description, package or remote endpoint, version.
• Run mcp-publisher login github, then mcp-publisher publish.
• Confirm the entry appears in the GitHub MCP Registry, then search @mcp <product> in VS Code's Extensions view.
• Add re-publishing to your release checklist. Downstream consumers carry your metadata only if it's current.

The trap: Claude Code's /mcp does not browse this registry. Checklist 3 is a separate, parallel placement, not a duplicate.

Checklist 2 · skills.sh (no submission step)

Any public GitHub repo containing SKILL.md files becomes installable via npx skills add owner/repo; the leaderboard auto-populates from install telemetry.

• Make the skills repo public with SKILL.md files. Repo quality is Play 4.
• Verify installability: npx skills add <owner>/<repo> exits 0 in a clean container.
• Confirm the listing auto-populated on skills.sh. There is nothing to submit.
• Embed the badge in your README: https://skills.sh/b/owner/repo.
• Point your own funnels at the install command: docs, README, launch posts. Installs are driven by your funnels, not directory browsing; the leaderboard is the scoreboard, not the channel.
• Log install counts as order-of-magnitude trend lines only. Telemetry is cache-noisy.

Checklist 3 · Anthropic Connectors + plugin marketplace (file first)

Reviewed surfaces, queried in-product: the Connectors Directory in Claude (~439 reviewed connectors) and claude-plugins-official, pre-loaded in Claude Code's /plugin Discover view.

• Meet the security and annotation requirements before submitting.
• Submit to the Connectors Directory at claude.com/connectors.
• Prepare the Claude Code plugin listing: .claude-plugin/marketplace.json.
• Submit the plugin marketplace listing (claude-plugins-official).
• Budget days–weeks for human review. File today, don't wait on the other checklists.
• After approval, verify in-product: Claude's connectors search and Claude Code's /plugin Discover view.
• Diarize renewal per Anthropic's listing policy (plugins are versioned).

Checklist 4 · Context7 (~1 hour + curation)

The one directory with a quality loop you control, plus real, measured demand for docs retrieval (~1.14M weekly npm downloads of its client).

• Run your 5–10 target queries on context7.com first and record who wins today (worksheet below).
• Submit the docs site, not the GitHub repo, at context7.com/add-library. Site entries outscore repo entries (Convex: 91.6 site vs 79.9 repo, an 11.7-point spread from source choice alone).
• Claim ownership of the entry (org/domain signals).
• Set the parse config: branch, include/exclude folders. Exclude changelogs, blog, and marketing pages.
• 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.
• Rewrite the description to contain your target terms, naturally phrased (worksheet below).
• Set a re-parse cadence tied to docs releases. Freshness correlates with benchmark score (Spearman −0.54; freshest five entries average 83.6 vs 72.3 for the stalest five).
• Record your baseline in the tracking grid: query positions, benchmark score, last-parse timestamp.

Description-engineering worksheet

The model, verified 2026-06-11: on the query "payments", DodoPayments ranks #1 (11,647 snippets, benchmark 83) while Stripe (265,284 snippets, benchmark 84.5) is absent from that query's top-10, because Stripe's entry is named simply "stripe." Smaller player, better description-to-query fit, category win. Task-phrase lexical space is largely unclaimed. The first product to put the phrase in its description owns the shelf.

Enumerate the 5–10 category nouns and verb phrases for your space; pick the top five and fill in:

Query 1: ____________________  Winner today: ________________  Your position: ____
Query 2: ____________________  Winner today: ________________  Your position: ____
Query 3: ____________________  Winner today: ________________  Your position: ____
Query 4: ____________________  Winner today: ________________  Your position: ____
Query 5: ____________________  Winner today: ________________  Your position: ____

Draft description (must contain the exact terms above, naturally phrased,
not stuffed):
____________________________________________________________________________
____________________________________________________________________________

Check: □ every target term appears   □ reads naturally   □ states tasks, not marketing

Tracking grid (weekly)

Re-check weekly; alert on any FAIL or two consecutive WARNs. Context7 thresholds: PASS = top-3 on the primary term; WARN = ≥2-position weekly drop; FAIL = benchmark below 80 or staleness over 14 days.

The five automated checks behind the grid (cron or CI, results appended to a dated CSV):

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.
4. Context7 position: rank for each of your five target queries, plus benchmark score and last-parse timestamp, scored against the thresholds above.
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.

The skip list

Skip these, or cap at five minutes for the backlink, and write the decision down so nobody "fixes" it later.

Tier 3 skipped on (date): ____________   Decision recorded by: ____________
Exception (if any) and why: ________________________________________________

Last verified 2026-06-11. Tiers, steps, URLs, and figures trace to the parent guide.

Parent guide: AI Agent Registries & Directories: The 4 That Matter · Hub: The Complete Playbook to Agentic Discovery