MCP server landscape
How 69,078 MCP servers are hosted, packaged and versioned: remote vs local, package registries, transports, secret declarations, categories and namespaces.
Executive summary
69,078 active MCP servers are indexed; 41,713 (60.4%) are remote-only, 20,838 local-only and 4,830 offer both.
Of 25,668 servers that ship a package, 17,679 use npm and 5,347 use PyPI; 25,344 packaged servers use the stdio transport.
9,988 servers (38.9% of packaged servers) declare at least one secret environment variable such as an API key.
Only 7,968 of 46,543 servers with a remote endpoint (17.1%) mark it as requiring authentication.
The registry holds 274,190 published versions; the median server has 1 and 37,793 servers (55.5%) have never published a second.
37,232 servers (54.7%) use a GitHub-verified io.github.* namespace; the rest use 24,065 DNS-verified domain namespaces.
Every figure above is computed from the database DER
Methodology
How this report was produced, and its limits
Generated on 2026-11-21 from the AgentSidekick database. The latest completed ingestion step finished at 2026-11-21 13:23 UTC. Tables queried: products, product_remotes, product_packages, product_versions, product_signals, security_findings, product_taxonomy, product_integrations, taxonomy_nodes, integrations. The catalogue holds 73,820 active products, of which 9,469 meet the data-completeness bar and have a profile page.
Classification method: product type comes from the source (registry entries are MCP servers by definition) or is inferred from repository topics; categories, capabilities, integrations and industries come from keyword rules in the editable taxonomy, so they are derived, not checked by a person. Curated records were researched against the official sources listed on each record.
Known limits:
- All figures describe what publishers declared in the official MCP Registry (server.json). Nothing here was installed or run.
- "Requires secret" means the package declares an environment variable flagged isSecret; it is a configuration signal, not a security assessment. A server can still need credentials it does not declare.
- "Requires authentication" on a remote endpoint is only counted when the registry entry says so; unmarked endpoints may still require auth.
- Launch dates are the first publish date in the source registry (official MCP Registry) or the GitHub repository creation date. They are not product announcements, and a re-published entry can look like a new launch.
- Categories, capabilities and integrations are assigned by keyword rules over names and descriptions (derived). Products that do not name a capability are missed, and generic words can produce false matches.
- Only 1 daily metrics snapshot(s) exist in product_metrics, so growth is measured from launch counts, not from changes in stars or downloads.
Note. Claim labels: source = stated by the publisher or registry; curated = checked by us against cited sources; derived = computed by a documented rule; heuristic = a rule of thumb with no measurement behind it. Full method: methodology.
Remote, local or both
The registry now lists more remote MCP servers than local ones: 41,713 remote-only against 20,838 local-only, with 4,830 offering both. A remote server is a URL the client connects to; a local one is a package the user installs and runs as a process on their machine. The shift matters for buyers because remote servers move the trust question from "what code am I running" to "who operates this endpoint and what data do they see".
MCP servers by hosting modelDER
Unit: servers. Top 4 shown.
Show the data as a table
| servers | |
|---|---|
| remote | 41,713 |
| local | 20,838 |
| both | 4,830 |
| unknown | 1,697 |
Authentication is rarely declared. 46,543 servers list a remote endpoint, but only 7,968 mark any endpoint as requiring authentication. That does not mean the rest are open: many registry entries simply omit the field. It does mean a client cannot rely on registry metadata to warn users before connecting.
Packages and transports
25,668 servers ship at least one installable package. npm is the default (17,679 servers), followed by PyPI (5,347); container images (OCI) and the MCP bundle format (mcpb) are smaller but established. stdio, where the client launches the server as a child process, is used by 25,344 packaged servers. The newer streamable HTTP transport appears on 45,804 remote servers and 347 packages, while the older SSE transport survives on 1,028 remote servers.
Package registries used by MCP serversSRC
| Registry | Servers | Packages |
|---|---|---|
| npm | 17,679 | 17,686 |
| pypi | 5,347 | 5,350 |
| oci | 1,971 | 1,983 |
| mcpb | 1,828 | 2,465 |
| nuget | 185 | 185 |
| cargo | 153 | 153 |
Transports declaredSRC
| Transport | Servers via package | Servers via remote endpoint | Remote, auth required |
|---|---|---|---|
| streamable-http | 347 | 45,804 | 7,530 |
| stdio | 25,344 | 0 | 0 |
| sse | 21 | 1,028 | 494 |
Secrets and configuration
9,988 servers declare at least one secret environment variable, 38.9% of the servers that ship a package. In practice these are API keys and tokens for the service the server wraps. For anyone deploying MCP servers to a team this is the configuration surface to manage: each such server is one more credential stored on a developer machine or in a client config file.
Note. A declared secret is a signal about configuration, not a safety verdict. Servers that need credentials but do not declare them are not counted.
Separately, the GitHub Advisory Database lists 92 published advisories against the npm or PyPI packages of 39 MCP servers (0.2% of packaged servers). An advisory is a verified fact about a package version; it does not mean the version a user installs today is affected, and the absence of an advisory is not evidence of safety.
Known advisories against MCP server packages, by severitySRC
Advisories may cover versions that have since been patched; check the first patched version on each product page.
| Severity | Advisories | Servers affected |
|---|---|---|
| critical | 13 | 8 |
| high | 46 | 26 |
| moderate | 24 | 13 |
| low | 9 | 6 |
What MCP servers connect to
By keyword classification, the broad categories below are the most common among MCP servers. 37,954 servers (54.9%) matched no category at all, usually because their description is a single generic line. 5,799 servers (8.4%) name at least one tracked third-party product.
Top categories among MCP servers (level-2 taxonomy, keyword-derived)DER
Unit: servers. Top 15 shown.
Show the data as a table
| servers | |
|---|---|
| Agent Commerce | 3,544 |
| Communication & Messaging | 2,521 |
| AI Coding Agents | 2,414 |
| Public & Real-world Data | 1,988 |
| Social & Streaming Platforms | 1,912 |
| Marketing | 1,839 |
| Crypto & Blockchain | 1,589 |
| Finance & Accounting | 1,583 |
| Cloud Providers | 1,562 |
| Files & Documents | 1,498 |
| Retrieval & RAG | 1,419 |
| Web Data | 1,410 |
| Agent Security | 1,261 |
| Agent Runtime & Deployment | 1,258 |
| Markets & Trading | 1,198 |
Most-mentioned integrations among MCP serversDER
Unit: servers. Top 15 shown.
Show the data as a table
| servers | |
|---|---|
| YouTube | 549 |
| GitHub | 423 |
| 403 | |
| Solana | 355 |
| 275 | |
| PostgreSQL | 257 |
| Shopify | 221 |
| Android | 211 |
| 189 | |
| Telegram | 186 |
| Stripe | 181 |
| Google Ads | 151 |
| WordPress | 149 |
| Google Maps | 117 |
| AWS | 115 |
What this does not show: a mention is not an implementation. A server that "works with YouTube" may expose one read-only endpoint or a full API. Category counts also double-count servers that match several categories.
Release cadence
The registry records 274,190 published versions across 68,059 servers. The median server has 1 version(s); 37,793 have only one and 4,442 have more than ten. 31,090 servers published a version in the 30 days before this build. A single version does not mean abandoned (some servers are thin wrappers that rarely change), but combined with no repository activity it is a reason to check before relying on one.
Servers by number of published versionsSRC
| Versions | Servers | Share |
|---|---|---|
| 1 | 37,793 | 55.5% |
| 2–3 | 16,607 | 24.4% |
| 4–10 | 9,217 | 13.5% |
| 11–50 | 4,016 | 5.9% |
| 51+ | 426 | 0.6% |
Versions published per monthSRC
Unit: versions per month, 2025-12 to 2026-11.
Show the data as a table
| Versions (versions) | |
|---|---|
| 2025-12 | 781 |
| 2026-01 | 1,174 |
| 2026-02 | 2,637 |
| 2026-03 | 5,724 |
| 2026-04 | 7,526 |
| 2026-05 | 9,112 |
| 2026-06 | 11,963 |
| 2026-07 | 18,067 |
| 2026-08 | 23,152 |
| 2026-09 | 39,308 |
| 2026-10 | 74,166 |
| 2026-11 | 78,416 |
Namespaces: who vouches for the publisher
Every registry entry lives in a namespace whose ownership the registry checks before publishing. 37,232 servers use io.github.* namespaces, verified by GitHub login, spread over 17,966 accounts. 30,827 use reverse-DNS domain namespaces, verified through DNS or HTTP, spread over 24,065 domains. Domain verification says the publisher controls a domain; it says nothing about the code.
Largest domain-verified namespacesSRC
Namespace = the part of the registry name before the slash.
| Namespace | Servers |
|---|---|
| io.usefulapi | 1,095 |
| dev.cotlet | 398 |
| ai.smithery | 209 |
| com.mambabuilt | 148 |
| software.brainiac | 147 |
| org.aimlds | 147 |
| com.textforagents | 135 |
| com.a2awire | 135 |
| app.wishpool | 125 |
| ai.getvda | 95 |
The 10 largest domain namespaces publish 2,634 servers between them, 8.5% of all domain-namespaced entries. Bulk publishing of generated wrappers explains much of the registry's growth, which is why profile pages here are limited to the 5,452 servers with complete records.
Recommendations
Editorial: the judgement of AgentSidekick, drawn from the figures above. Not a measurement, not investment advice.
- Editorial
client vendors should surface the authentication field. Only 17.1% of remote servers declare it, so users connect without knowing whether an endpoint is open.
- Editorial
clients that install packaged servers should check advisories at install time; 39 servers already have at least one published advisory against their package.
- Editorial
server authors should declare secrets explicitly. 9,988 servers already do; the rest leave users to discover required keys at runtime.
- Editorial
new servers should target streamable HTTP rather than SSE for remote use; SSE remains on 1,028 remote servers against 45,804 on streamable HTTP.
- Editorial
when choosing a server, prefer one with more than one published version and an active repository; 55.5% of servers have only ever published once.
Sources
- Official MCP Registry: registry.modelcontextprotocol.io (external)checked
- GitHub repository metadata: docs.github.com (external)checked
- MCP Registry documentation: github.com (external)checked
- GitHub Advisory Database: github.com (external)checked
- MCP specification: transports: modelcontextprotocol.io (external)checked
- AgentSidekick methodology: agentsidekick.dev (external)checked
Data behind this report
- Index of profiled products: /api/dataset/products.jsonJSON, regenerated each build
Cite this report
MCP server landscape. AgentSidekick, 2026-11-21. https://www.agentsidekick.dev/reports/mcp-server-landscape/
Figures change when the database is refreshed; cite the date shown.