Guide Jun 10, 2026

How We Publish MCP Servers and Claude Skills to Every Marketplace — Our 2026 Playbook

How to Publish MCP Servers & Claude Skills in 2026

Most builders push their MCP servers and Skills to GitHub and wonder why nobody finds them. After shipping two of our own across 13 catalogs, here's the exact playbook we use — what's automated, what's manual, and where most authors lose a week.

TL;DR

Why this matters

Building MCP servers and Claude Skills is the easy part. Getting them in front of the 2.5M+ Claude users is the hard part — and most builders skip it entirely, push to GitHub, and wonder why nobody finds their work.

We at NeCL publish open-source tooling as a way to earn inbound for our paid AI engineering work. Every MCP we ship is also a marketing surface. We measured this directly: our last two opensource releases (a Telegram bot framework and an HN MCP) brought in 7 inbound Calendly requests in the first month, all tagged with utm_source=marketplace. The catalogs distribute the artifact; our README converts the curious into Calendly bookings.

This is the exact playbook we run internally, sanitized for anyone who wants to follow it. We've shipped four open-source artifacts this quarter using these same steps.

What you'll need

The 13 catalogs (in order of ROI)

For MCP servers (5 main + community lists)

  1. mcp.so — biggest aggregator (20,222+ servers). Submit URL form. Indexes downstream from awesome-mcp-servers. Highest impression volume.
  2. registry.modelcontextprotocol.io — official Anthropic registry. JSON manifest via GitHub PR. Maximum credibility, 1-7 days review.
  3. glama.ai/mcp — curated catalog with use-case focus. Submission form. 2-5 day review.
  4. awesome-mcp-servers (github.com/punkpeye/awesome-mcp-servers) — community awesome-list. PR with a single line. Feeds downstream catalogs.
  5. Smithery.ai — was the CLI publisher, now hosted-only. Skip for stdio servers unless you're ready to deploy to Railway/Fly.io. Use .mcpb bundle instead (see below).

For Claude Skills (5 main)

  1. skillsmp.com — main aggregator (658+ skills). Auto-indexes public repos with SKILL.md. Submit URL to skip the queue.
  2. claudeskills.info — open-source-only catalog. Submit form or auto-index.
  3. claudemarketplaces.com — multi-type directory (Skills + MCP + plugins). Submit form, 1-7 days review.
  4. claudepluginhub.com — lets you create your own branded sub-marketplace. Major SEO win — branded URL like claudepluginhub.com/yourname.
  5. daymade/claude-code-skills — curated awesome-list. PR with a single line.

The Smithery problem

In 2026, Smithery deprecated stdio MCP publishing. The CLI now requires either: - A live HTTP/SSE server URL (you host it on Cloudflare/Railway/Fly.io) - An .mcpb bundle file from Anthropic's mcpb CLI

For most open-source MCP authors, the .mcpb bundle is the right answer:

npm install -g @anthropic-ai/mcpb
cd your-mcp-repo
mcpb init   # generates manifest.json
mcpb pack   # creates .mcpb bundle

Drop the bundle into Claude Desktop → Settings → Developer for instant install. Attach it to a GitHub Release for download links.

Repo strategy — monorepo vs separate repos

This matters more than people think. Each catalog indexes differently:

Catalog Indexes by Monorepo result
skillsmp.com, claudeskills.info per SKILL.md file ✅ Multiple entries from one repo
Smithery (Skills) per SKILL.md folder URL ✅ Multiple entries
mcp.so, glama.ai per repo URL ❌ One entry per repo — monorepo hides skills
awesome-listы manual lines ✅ You add as many lines as you want

Conclusion: - Skills: one monorepo with multiple SKILL.md folders. Catalogs parse each one individually. - MCP servers: one repo per server. Monorepo MCPs get one combined listing on mcp.so and lose visibility.

The 5-step launch sequence

  1. Prep (1 hour): Write your README with Built by [Your Brand] everywhere. Add a Need a custom one? section with Calendly link.
  2. Push to public GitHub (15 min): Two repos — your-skills/ and your-mcp-name/. Add topics for GitHub Search discoverability.
  3. Submit forms (30 min): Hit mcp.so, glama.ai, claudemarketplaces.com, skillsmp.com, claudeskills.info — one tab each.
  4. Open PRs (45 min): awesome-mcp-servers, daymade/claude-code-skills, registry.modelcontextprotocol.io. Each is a 5-line PR.
  5. Wait 1-2 weeks for indexers to cascade. Downstream AI-search catalogs and Google will surface you organically after the cascade settles.

What we shipped

At NeCL we just published two pilot pieces with this exact playbook:

Both are MIT-licensed. Both built from production patterns we already use internally. Both have the brand attribution wired through every catalog.

See also

FAQ

Do I need to be on PyPI or npm to publish an MCP server?

No. Use `uvx --from git+https://github.com/you/repo.git your-package` in the install instructions. Works without any package registry publish.

Is Smithery still worth it for stdio MCPs?

Not unless you deploy your server publicly. For local-only stdio MCPs, an `.mcpb` bundle gets you the same Claude Desktop one-click install without the hosting bill.

How long until catalogs actually surface my work?

Forms and PRs: 1-7 days for human review. Cascade indexers (mcp.so picking up from awesome-list, AI search picking up from Google): 1-3 weeks for first surface.

Will publishing open source actually drive paid work?

For us, yes — every README ends with a Calendly link tagged `?utm_source=marketplace`. We treat catalog impressions as cold outreach with brand context already established.

Can I publish private MCP servers and Skills?

No. All public catalogs require public GitHub repos. For client-private work, you'd run your own internal catalog or share the bundle file directly.

What about Cursor, Continue, Cody marketplaces?

They have their own extension marketplaces that don't index public GitHub. You'd submit there separately. We cover those in a future post.

Need help building or distributing your own MCP server or Claude Skill? We do this work for hire — from architecture to publishing across all 13 catalogs. .

Book a 30-min call