MCP Server Marketplaces Compared: What Actually Runs

Paste a Streamable HTTP endpoint into Claude's connector config and save it. If the tools appear and a call returns data, the listing was real. If the directory handed you only an npx install line, nothing will paste at all. That gap is the point of this comparison. https://mcp.moltlinestudio.com/<server> is a pasteable URL; a stdio command is not.

MCP defines two standard transports: stdio, where the client launches the server as a subprocess, and Streamable HTTP, where the client connects to a URL. The spec says clients SHOULD support stdio whenever possible, so a stdio-only listing is not broken. It just cannot pass a paste-a-URL test in Claude, Cursor or Codex.

This post compares MCP server marketplaces and directories on that practical question: does the listing lead to a tool that runs today, and what does it cost to start? I cover the five criteria in the comparison table and the exact test procedure for each client. Later sections handle free-tier limits, the 2026-07-28 spec changes around stateless cores and OAuth 2.1, and where Moltline Studio fits: 22 hosted servers, 160 tools, 110 free without signup.

Can You Paste an MCP Endpoint Into Claude and Run a Tool Today?

Run this first:

curl -i https://moltlinestudio.com/api

It returns HTTP/2 402 with an x402 payment challenge in the body, which is how an agent discovers a price with no human in the loop.

Every directory promises discovery; a listing is not a connection. Paste the URL into your client and watch whether a tool call returns.

That is the only test used here: the endpoint pastes into Claude, Claude Code, Cursor or Codex CLI and a tool call succeeds on the day you read this. Nothing else on a directory page counts.

Transport details were checked against the 2025-03-26 spec. The 2026-07-28 spec revision changes several of those rules. This comparison has a shelf life.

Keep this command in view:

claude mcp add --transport http <name> https://mcp.moltlinestudio.com/<server>

Swap in a real slug from /servers.html. For a worked walkthrough, see pasting an MCP endpoint into Claude Code.

What Separates a Usable MCP Listing From a Dead One?

Five rows decide whether a listing is usable today. Score each source against them before wiring anything in.

1. Transport shown. The spec defines two standard transports: stdio, where the client launches the server as a subprocess, and Streamable HTTP (transports spec). A listing printing only an npx or uvx install command is stdio. There is no URL to paste.

2. Stdio is not disqualifying. Clients SHOULD support stdio whenever possible. Stdio listings are runnable, not pasteable. Judge them on hosting and maintenance instead.

3. Auth required. The 2026-07-28 spec revision adds OAuth 2.1 plus OIDC as a requirement, per its migration notes. If a listing is silent on auth, you meet the flow at the first tool call. Silence is not "none."

4. Cost to start. Four things get called free: no signup, free behind an account, free code you host and pay to run, and a free tier with a paid unlock. Note which one applies before you commit. That distinction shapes what to look for when evaluating an MCP server.

5. Public evidence. Can you open a result page for the exact server? A grade, a repo, a live endpoint. If the only proof is a marketing line, that is the proof.

Cut rule: if a row cannot be verified against the source's own page, cut the row. Never guess its pricing or features.

The Comparison Table: Five Criteria, One Row Per Source Type

Score each source against the five criteria above, using only what its own page states.

Checked 22 July 2026 against the 2025-03-26 spec. "Not stated" means the page is silent; unstated auth or price hides a rejection or a bill. Re-check every row against the 2026-07-28 spec revision.

How Do You Test a Listing in Claude, Cursor and Codex?

The table tells you which listings to try. The test is the same everywhere: paste the URL, make one tool call.

Claude Code. Run:

claude mcp add --transport http moltline https://mcp.moltlinestudio.com/<server>
claude mcp list

Use a real slug. If your build rejects --transport http, the endpoint will not work. See the 60-second Claude Code walkthrough.

Cursor. Add a URL entry to ~/.cursor/mcp.json, then reload:

{"mcpServers": {"moltline": {"url": "https://mcp.moltlinestudio.com/<server>"}}}

A stdio-only listing has nothing for that url field. That is the test.

Codex CLI. Register the same URL in your Codex config. If your build only accepts command and args, the listing cannot be pasted. It fails for you even if the server is healthy.

No client? POST a JSON-RPC initialize to the endpoint and read the status. Log these:

• 404: wrong slug or no server at that path.

• 401 or 403: auth required.

• 402: payment required.

A 402 is not downtime. On moltlinestudio.com/api it is the live x402 challenge: the server states the price and waits for on-chain settlement. Treat it as a working connection with a gate.

A connection reset means HTTP+SSE only, the older transport that Streamable HTTP replaced in the 2025-03-26 spec. That is the operator's migration item, not a client-side fix.

What Does It Cost to Start Using an MCP Directory?

Cost to start is the row readers skip and then regret. Four cases, ordered by real cost:

1. No signup, no payment. $0, paste an endpoint and call a tool.

2. An account you create. $0, plus key rotation and a signup you did not want.

3. A free install you host yourself. A container, a runtime, upgrades and a process to keep alive. Budget maintenance time before calling it cheap.

4. A paid licence. $19/month, paid in cryptocurrency through NOWPayments (ETH, USDT, SOL and 300+ coins).

Our numbers, stated plainly: 22 hosted servers, 160 tools, and 110 free forever with no account, no API key and no signup. The other 50 unlock with one $19/month All-Access licence. The getting-started guide covers the first paste.

No account means no usage dashboard, no seat management, no admin console. Support is one person reading an inbox. That is the trade for not making you sign up.

Agent-driven purchase is unusual. https://moltlinestudio.com/api serves a live HTTP 402 x402 challenge, so an agent discovers the price and settles on-chain with no human approval (x402). A human still funds the wallet.

Where Moltline Studio Fits in This Comparison

Moltline Studio is not a directory you browse. It runs 22 hosted MCP servers exposing 160 tools, each with its own Streamable HTTP endpoint:

claude mcp add --transport http moltline https://mcp.moltlinestudio.com/<server>

Swap <server> for a slug from the hosted server list. One hundred ten of those tools are free with no signup and no API key.

It is also not a marketplace. Only tools this one person builds and runs are hosted here. If you need third-party servers in a single catalog, this is the wrong source. That limitation belongs next to the facts, not in a footnote. The multi-server MCP architecture notes explain how the set fits together.

Check these before trusting any of the above:

• 138 SKILL.md skills open at github.com/GarphenGate/moltline-oss, plus a free SKILL.md linter

• A free agent-readiness checker at moltlinestudio.com/agent-check.html, with 21 checks and no signup

• MCPize audit grades running A+ to B+, each with its own public result URL

No grade gets printed here without its URL. Bundles also ship through Agensi, so some of this material is reachable without visiting moltlinestudio.com at all.

What the Free Tier Cannot Do

The free tier is the honest starting point, so here is what it does not cover.

• The 50 premium tools sit outside it. There is no trial unlock. The cheapest way in is a 2 USDC day pass bought over x402 at moltlinestudio.com/api/premium/day-pass, valid to the end of the next UTC day.

• Payment is a hard gate for those 50. The HTTP 402 x402 challenge lets an agent pay without a human in the loop, but someone still has to fund the wallet once.

• Endpoints are remote Streamable HTTP only. A client that speaks stdio needs one that supports Streamable HTTP. There is no local npx install to fall back on. That is a real constraint for older setups.

• No account means no SSO, no audit trail and no admin console. Teams get five seats on one invoice through the $99/month Team plan, arranged by email. If your organisation needs SSO or an audit trail, this source is not the fit, whatever the tool count says.

• Spec risk, on the record: endpoints were tested against the 2025-03-26 spec. The 2026-07-28 spec revision adds OAuth 2.1 and OIDC requirements, per its migration notes. Re-run the paste test against it and expect auth flows to change.

Run the paste test on a premium endpoint before you buy the licence.

What Is Changing in the MCP Spec, and Does It Break Your Setup?

That spec risk is worth spelling out. Everything below comes from the aaif.io migration post on the 2026-07-28 spec revision, so check it against the spec text itself before planning a migration.

Three changes it describes hit endpoints you already pasted:

• Stateless core. The revision targets plain round-robin load balancing, dropping sticky sessions and shared session stores. The post calls those a bottleneck for deployed remote servers.

• HTTP+SSE listed for deprecation. An SSE-only listing is a migration item. Your paste test can pass on one spec version and fail on the next.

• Roots, sampling and logging listed for deprecation, with OAuth 2.1 plus OIDC added as a requirement. Premium and authenticated endpoints carry the most exposure, since they are the ones running auth flows.

The post also describes a Tasks extension, returning a durable taskId and polling through tasks/get. A listing that claims long-running jobs without saying how it polls is unverified.

Practical rule: when your client or server moves to the 2026-07-28 revision, re-run the paste test on every endpoint you depend on. A directory listing that cannot state its spec version is not evidence of anything. Where to Find More covers sources that do state theirs.

Which MCP Source Should You Use?

Spec version decides whether an endpoint keeps working across revisions. Source choice decides whether it works in the next five minutes. Match the source to the job:

• You need a URL to paste now. Use a source that publishes remote Streamable HTTP endpoints. A listing that only prints an npx or uvx install command has no URL, so it fails the test, however good the server is.

• You need zero signup and zero cost to evaluate. Moltline Studio: 110 of 160 tools are free forever, with no account and no API key. Cheapest honest start here. Try a free server page.

• You need third-party servers in one catalog. Use a marketplace. This studio is not one. Only tools built and run here sit behind mcp.moltlinestudio.com. Cheaper to say than to discover after signup.

• You need audited evidence before production. Open the MCPize result URL for that specific server. Grades run A+ to B+, each page public.

• You need source or agent skills, not just endpoints. github.com/GarphenGate/moltline-oss holds 138 SKILL.md files. Agensi carries the bundles.

• You are stuck with a stdio-only client. A stdio listing is your only option. Price the container, runtime and upgrades before you call it free.

What Should You Do After Comparing MCP Directories?

Pick one row from the table and run it. Before opening another directory, add a real endpoint to a real client:

claude mcp add --transport http moltline https://mcp.moltlinestudio.com/<server>

Swap <server> for a slug from the server list, then make one tool call. Three outcomes:

• It works. You have a source.

• It returns a premium_tool_requires_license error in the tool result. The tool is premium. The error names the $19 licence and the 2 USDC day pass, and points your agent at the x402 challenges on moltlinestudio.com.

• It fails. The listing was discovery, not a connection.

Then write the five criteria into your own notes: transport shown, pasteable endpoint, auth required, cost to start, public evidence. Score every directory against them. Cut anything that cannot answer all five.

Re-test against the 2026-07-28 spec revision. Its OAuth 2.1 and OIDC requirements plus the HTTP+SSE deprecation change which listings still work.

One working endpoint beats a bookmarked directory. Keep the comparison table next to your config file.

Conclusion

The comparison comes down to four things. A usable MCP source shows its transport, gives you a pasteable endpoint, states auth and cost upfront, and publishes auditable evidence. Anything less is discovery, not a connection.

So skip the bookmarking. Pick one row and prove it works:

• Choose a slug from the server list.

• Run claude mcp add --transport http moltline https://mcp.moltlinestudio.com/<server>.

• Make one tool call. Read the result.

Works, and you have a source. Returns a licence error on a premium tool, and you have a price your agent can settle at moltlinestudio.com/api. Fails, and you learned that cheaply.

Score every directory against those five criteria. Re-test against the 2026-07-28 spec revision. One live endpoint beats a hundred bookmarked listings. Go make the call.