Original research · 2026 Q4 · running edition, updated on every publish
State of the MCP Ecosystem, 2026 Q4
Every server in the official MCP registry, measured: advisories that affect the version it ships, unpinned installs, remote-only and source-less servers, abandoned repositories and possible typosquats. Generated from MCP Server Watch data, reproducible from the downloads below.
Data as of 2026-11-22 13:15 UTC · Quarter 2026-10-01 to 2026-12-31 · CSV · JSON · Methodology · Editions
Key findings
- 61.3% of the 68,913 servers in the official MCP registry are remote-only: they publish a hosted endpoint and no package, so there is no code to review or pin.
- 36.7% link no source repository. Of the 43,502 that do, 18,603 had a commit in the last 30 days and 290 are archived.
- 10 servers ship a package version with a published advisory (18 distinct advisories); 4 have at least one with no fixed version.
- OSV lists 5 malicious-package records against packages that registry servers ship (any version), among 215 advisories on 66 packages.
- Beyond the registry, MCP Server Watch tracks 1,378 more servers found on GitHub (658), npm (465), Docker Hub (212), PyPI (43); they are not in these tables.
- The 10 largest publishers account for 16.3% of registry entries.
- 25,285 servers have a current version published in 2026-11 so far (16,283 in 2026-10).
Advisories affecting the shipped version
A server counts once, under its most severe advisory.
Show table
| Group | Servers | Share |
|---|---|---|
| Critical | 1 | 10.0% |
| High | 7 | 70.0% |
| Medium | 1 | 10.0% |
| Low | 1 | 10.0% |
Signals are facts, not a verdict. A server can carry several.
Show table
| Group | Servers | Share |
|---|---|---|
| Remote endpoint only | 42,256 | 61.3% |
| No source repository | 25,311 | 36.7% |
| Repository facts unavailable | 8,595 | 12.5% |
| Archived repository | 290 | 0.4% |
| No commit in 12 months | 180 | 0.3% |
| Unpinned install | 50 | 0.1% |
| Advisory | Package | Severity | Servers | Fixed in |
|---|---|---|---|---|
| GHSA-2m9m-6cr5-9x25mssql-mcp-core: SQL injection in mssql-mcp write tools via unparameterized table names and WHERE clauses | npm:@connorbritain/mssql-mcp-reader | High | 3 | 0.5.0 |
| GHSA-p34x-fmph-9fjxFlyto2 Core: Arbitrary file write via unguarded data.*/file.* modules (incomplete fix of CVE-2026-67429) | pypi:flyto-core | Critical | 1 | 2.26.8 |
| GHSA-jx74-cqjv-2c67Flyto2 Core: Unauthenticated flyto-verification /run: callback_url SSRF and internal runner-secret exfiltration | pypi:flyto-core | Critical | 1 | 2.26.7 |
| GHSA-2956-977x-2w3rFlyto2 Core: Arbitrary file write via image.download (and other file-writing modules) | pypi:flyto-core | Critical | 1 | 2.26.7 |
| GHSA-mx6p-3fg7-v6pjobsidian-sync-mcp: Missing Host/Origin validation lets a public website drive a victim's Obsidian vault via DNS rebinding (no-auth mode) | npm:obsidian-sync-mcp | High | 1 | 0.5.7 |
| GHSA-fq6x-j3g6-xx83@rog0x/mcp-docker-tools: executes arbitrary code via the handleImageList function | npm:@rog0x/mcp-docker-tools | High | 1 | no fix listed |
| GHSA-2vq3-mpmf-9xgr@justice8096/sagemath-mcp-server: executes arbitrary code via the escapeShellString and sanitizeCode functions | npm:@justice8096/sagemath-mcp-server | High | 1 | no fix listed |
| GHSA-jwc4-922r-f2rw@rog0x/mcp-git-tools executes arbitrary code via the resolveBaseBranch function | npm:@rog0x/mcp-git-tools | High | 1 | no fix listed |
| GHSA-3r68-x3xc-rxpgwireshark-mcp vulnerable to arbitrary file write via export_objects when WIRESHARK_MCP_ALLOWED_DIRS is not configured | pypi:wireshark-mcp | Medium | 1 | no fix listed |
| GHSA-x3cv-r3g3-fpg9Neo4j Labs MCP Servers: SSRF and Data Modification via read_only Mode Bypass Through CALL Procedures | pypi:mcp-neo4j-cypher | Low | 1 | 0.6.0 |
How servers are distributed
A server can ship packages to several registries.
Show table
| Group | Servers | Share |
|---|---|---|
| Remote only (no package) | 42,256 | 61.3% |
| npm | 17,860 | 25.9% |
| PyPI | 5,423 | 7.9% |
| OCI (Docker) | 1,990 | 2.9% |
| MCP bundle (mcpb) | 1,861 | 2.7% |
| No package or remote | 687 | 1.0% |
| NuGet | 186 | 0.3% |
| Cargo | 153 | 0.2% |
Show table
| Group | Servers | Share |
|---|---|---|
| Streamable HTTP | 46,411 | 98.4% |
| SSE (legacy) | 1,123 | 2.4% |
Popularity and maintenance
Show table
| Group | Servers | Share |
|---|---|---|
| No GitHub repository | 25,411 | 36.9% |
| Repository facts unavailable | 8,595 | 12.5% |
| 0 stars | 20,554 | 29.8% |
| 1–9 | 10,463 | 15.2% |
| 10–99 | 2,661 | 3.9% |
| 100–999 | 939 | 1.4% |
| 1k–9.9k | 230 | 0.3% |
| 10k or more | 60 | 0.1% |
Show table
| Group | Servers | Share |
|---|---|---|
| Commit in last 30 days | 18,603 | 27.0% |
| 31–90 days | 10,542 | 15.3% |
| 91–365 days | 5,582 | 8.1% |
| More than a year (stale) | 180 | 0.3% |
| Repository facts unavailable | 8,595 | 12.5% |
| No GitHub repository | 25,411 | 36.9% |
The current month is partial.
Show table
| Month | Servers |
|---|---|
| 25-12 | 192 |
| 26-01 | 276 |
| 26-02 | 761 |
| 26-03 | 1,399 |
| 26-04 | 1,508 |
| 26-05 | 1,945 |
| 26-06 | 2,225 |
| 26-07 | 3,418 |
| 26-08 | 5,480 |
| 26-09 | 9,442 |
| 26-10 | 16,283 |
| 26-11 | 25,285 |
Possible typosquats
1 foundLow-star npm/PyPI packages whose name is one character away from a popular registry package, or that take a popular scoped package's name without its scope. Name similarity is a lead for review, not evidence of intent: these may be forks, renames or unrelated projects. ThreatDeck has not confirmed any of them as malicious.
| Package | Resembles | Difference | Stars (package / popular) |
|---|---|---|---|
| npm:concord-mcp | npm:@concord-ai/concord-mcp | scope dropped | 2 / 1,005 |
AI Incident Ledger lists the OSV malicious-package records for registry packages with their sources.
Advisory speed
Time from an advisory's publication to its listing on this board (checked every 6 h), and how far the vendor or GitHub advisory came before NVD. Listing speed counts only advisories published after we started watching (2026-11-10); older advisories were back-filled. Median time to listing so far: 24.6 h over 2 advisories (90th percentile 44.8 h). Vendor/GHSA advisories came a median 18.1 h earlier than NVD.
| Tool | Tracked | Measured | Median listing | Before NVD | In NVD | Lead vs NVD |
|---|---|---|---|---|---|---|
| OpenClaw | 101 | 1 | 4.5 h | 1/1 | 1 | n/a |
| GitHub Copilot | 7 | 1 | 44.8 h | 0/1 | 7 | n/a |
| Claude Code | 32 | 0 | n/a | n/a | 31 | 18 h earlier |
| Cursor | 26 | 0 | n/a | n/a | 26 | 27.4 h earlier |
| Roo Code | 16 | 0 | n/a | n/a | 16 | 15.5 h earlier |
| Zed (AI agent) | 12 | 0 | n/a | n/a | 12 | 5.8 h earlier |
| OpenAI Codex CLI | 4 | 0 | n/a | n/a | 4 | 46.3 h earlier |
| Aider | 3 | 0 | n/a | n/a | 3 | 0.7 h later |
| Cline | 3 | 0 | n/a | n/a | 2 | 19.6 d earlier |
| Continue | 1 | 0 | n/a | n/a | 1 | n/a |
| Gemini CLI | 1 | 0 | n/a | n/a | 1 | 60.8 d earlier |
| goose | 1 | 0 | n/a | n/a | 1 | 17.1 d earlier |
| OpenHands | 1 | 0 | n/a | n/a | 1 | 3.3 d earlier |
Tracked = advisories listed on the board (newest 100 per tool). Lead vs NVD = median time by which the vendor or GitHub advisory came before NVD's publication. "In NVD" counts advisories whose CVE has an NVD publication date so far (looked up a few at a time). Advisories that come from NVD itself have no lead time.
Methodology
- Population: every server listed in the official MCP registry at the time of the crawl (68,913 servers, latest version of each). Servers that MCP Server Watch found elsewhere (npm, PyPI, GitHub, Docker Hub) are not in the tables; they are counted separately.
- Repository facts (stars, last commit, archived) come from the GitHub API for the repository each server declares; servers without a GitHub repository have none.
- Advisories come from OSV (GitHub Security Advisories, PyPI, OSSF malicious packages) for each npm/PyPI/NuGet/Cargo package a server ships. A server counts as affected only when an advisory's range includes the version the registry entry ships.
- Signals are factual: unpinned = a package in the registry entry has no version or a floating one ("latest", "*") (0.1%); remote-only = only a hosted endpoint, no package or source to review (61.3%); no source = the registry entry links no repository; stale = no commit in 12 months; archived = the repository is archived.
- Possible typosquats: npm/PyPI packages with fewer than 50 stars whose name is one character away from a registry package with 1,000+ stars, or that take a popular scoped package's bare name without its scope. The distinctive part of the popular name (generic tokens such as "mcp" and "server" removed) must be at least 5 characters; digit edits, the same publisher and the same repository are excluded.
- Months count servers by the publication date of the registry entry's current version (re-published servers move to the month of their latest release).
- Publisher concentration uses the registry namespace (the part before "/"); publisher names are not listed.
- The report is regenerated on every publish from the same dataset as MCP Server Watch; the quarter's edition is frozen when the quarter ends. The JSON and CSV downloads contain exactly the tables shown.
Limitations
- The official registry is a subset of the MCP servers developers run; many servers live only on npm, PyPI, Docker Hub or GitHub.
- OSV coverage is incomplete: no advisory does not mean no vulnerability, and remote-only servers cannot be checked at all.
- Advisory matching uses the version in the registry entry; installs that pin an older or newer version differ.
- Stars measure attention, not quality or safety. Repository facts can be up to a few days old.
- Name similarity is a lead for review, not evidence of intent: listed packages may be forks, renames or unrelated projects. None has been confirmed as malicious by ThreatDeck.
- Registry entries can be published in bulk by a few namespaces, which inflates totals for some categories.
Reproduce and cite
The JSON and CSV hold every table on this page (aggregates and advisory/package ids only; no publisher names). Per-server data is browsable in MCP Server Watch and through the API. Cite as: ThreatDeck, "State of the MCP Ecosystem, 2026 Q4", data as of 2026-11-22.
Editions
- 2026 Q4 running · generated 2026-11-22