The agent economy spent 2026 building the rails and has now started counting the sleepers. Between November 10 and November 16, three documents landed that together form the first real inspection of the Model Context Protocol ecosystem: a full census of every host behind the official MCP Registry, a hands-on audit of MCP security scanners, and a community roundup of spec changes, SDK releases, and security fixes. Read together, they describe infrastructure that grew faster than its hygiene, at the exact moment the first agent-assisted holiday shopping season gets underway.

The short version: 59,944 active servers in the registry. 16,470 hosts will list their tools to anyone who asks, with no sign-in. At least 15,064 tools sitting on those open lists describe themselves as destructive. Only 5.2 percent of version-reporting servers speak the spec revision that carries the current security model. And the scanners built to police all of this produce roughly 78 percent false positives at the pattern layer.

If your AI agent shops for you, this is the plumbing it drinks from. Here is what the census actually found, and what it means for commerce.

A registry of 59,944 claims, not 59,944 vettings

On November 10, 2026, at 06:07:30 UTC, the official MCP Registry held 59,944 active servers, counting the latest version of each. Of those, 40,486 (68 percent) offer a remote endpoint reached over HTTPS, 23,328 ship a package that runs on the user’s machine, 4,535 do both, and 665 offer neither. The remote entries point at 30,402 distinct hosts once you collapse duplicates, which is the population Spellwarden probed.

Publishing is concentrated at the top and spread thin everywhere else. Of 37,147 namespaces that published at least one server, 33,803 published exactly one. Meanwhile, 21 namespaces with 100 or more servers each account for 11,398 entries, 19 percent of the registry, and the largest single namespace holds 2,366. Fifty-seven percent of all entries sit under an io.github.* namespace, which the registry grants to whoever authenticates as that GitHub user or organization owner.

The report’s own framing deserves quoting in full, because it is the sentence every team connecting agents should internalize: “A listing is a statement of who published it, not a security review.” The registry is in preview, holds metadata only, and delegates security scanning to package registries and aggregators.

The open front door

Spellwarden sent one unauthenticated probe to each of the 30,402 hosts, making at most nine requests, exactly as an MCP client does on first connection. 29,277 hosts (96 percent) gave an HTTP answer. Of the 26,548 that behaved like MCP servers, 16,470, or 62 percent, listed their tools to a client that had never signed in. The median open list exposes six tools.

An open tool list is not automatically a flaw. Documentation search and public data services are meant to be open. The report puts the risk boundary precisely: “It becomes a risk when the tools write, delete or spend money and nothing asks who the caller is.”

That is not a hypothetical boundary. Of the open lists, 9,691 include at least one tool that does not declare itself read-only. At least 15,064 tools on open lists declare themselves destructive, and the count is a floor because some lists could not be read in full. Across the 196,542 tools on lists Spellwarden could read completely, 116,663 (59 percent) claim read-only status and 41,165 (21 percent) carry no hint either way. And those hints are self-reported: “The hints are the server’s own claim, so a read-only hint is a statement, not a guarantee.”

Two fairness notes the report makes about itself, and that critics should respect: the probe never called a tool, sent a credential, or tried to get past a sign-in, and a server can list tools openly while still checking the caller on every actual call. The census measures what an anonymous client can see, not what it can do. That distinction is exactly why the numbers matter: what an anonymous client can see includes the text your model will read and follow.

Text written to be obeyed

Tool descriptions are not documentation for humans. They are instructions the model reads and tends to obey. That is the entire mechanism of tool poisoning: hide directives in a description, and the agent carries them out.

Spellwarden’s pattern rule SW-MCP-011 fired on 967 hosts, 5.9 percent of hosts with an open tool list. On 932 of them, at least one description matched a concealment, override, or injected-instruction pattern. The team then drew a random sample of 60 flagged hosts and reviewed every flagged description in full. Nine only steered the model toward the tool: call it first, always use it, use it instead of answering directly. Six went further. One is unambiguous, and it is worth quoting exactly: it “tells the model to call the tool without asking and not to mention it to the user afterwards.”

Scaled to the flagged population, that is roughly 145 hosts whose only instruction-like text steers the model (95 percent interval 78 to 252) and roughly 96 whose descriptions ask it to hide something, follow a script, or collect data (45 to 194). Against all 16,435 open tool lists, that is about 0.9 percent steering and about 0.6 percent more serious, with about 16 clear-cut cases.

Those are floors, and Spellwarden says so. On November 13 the team published a recall measurement for its own scanner: version 1.12.0 catches 9.8 percent of instructions on unseen descriptions at 75 percent precision, and on a held-out set of 7,445 hosts, between 462 and about 950 carry an instruction of some kind while the rule flags 213. In the report’s words, counting what the rule misses, “the share is much higher, so read those figures as the floor they were.”

Anyone who has followed the fake review economy will recognize the shape of this problem instantly. Reviews are text written to persuade humans; tool descriptions are text written to steer machines. The economics of manipulation migrate to whichever surface the buyer, human or agent, actually reads. The difference is that a poisoned review inflates a star rating, while a poisoned tool description can decide what an agent does before the user’s own request gets a vote.

The false-positive wall

If the scale of the problem is census-level, the tooling to detect it is early. The AppSec Santa audit ran two open-source scanners, Cisco’s mcp-scanner and Invariant Labs’ MCP-Scan, across 33 local MCP servers covering 433 tools. The pattern scanner flagged 27 detections across 10 servers. After reviewing every one, the auditor classified only 6 as genuine security concerns. False positive rate: roughly 78 percent.

The failure modes are instructive. Eight “prompt injection” flags fired on standard dependency language like context7’s “You MUST call this function before ‘Query Documentation’ tool,” which is simply how well-documented MCP tools declare ordering. A browser tool got flagged for script injection because its description says it returns HTML “with script tags REMOVED,” which the audit calls “the opposite of injection.” A project management tool got flagged for PII exfiltration because the pattern matched the word “name.”

Meanwhile the six genuine findings show what real risk looks like: desktop-commander’s file search can locate .ssh and .aws directories, and combined with a code execution finding on the same server, the audit concludes it “has a real dual-threat profile: find credentials, then exfiltrate them.” A PostgreSQL server exposes arbitrary function creation and execution. A git server runs operations on user-controlled input.

The audit’s answer is a three-layer stack: pattern scanning (what text is present), config-drift scanning (whether tool definitions mutate between calls), and supply-chain auditing (auth posture, typosquats, auto-spawning configs). The third layer found 4 of 10 audited remote servers genuinely running open endpoints with no authentication, and flagged both audited project-config repositories for auto-spawning MCP servers the moment an agent trusts the folder. It also false-flagged an official MCP reference server as a typosquat. No layer is clean on its own, and stacking immature layers is expensive engineering.

The takeaway for anyone routing agents through third-party servers: automated detection is not yet a trust decision you can delegate. Curation still beats scanning.

The spec is outrunning the fleet

The protocol’s governance, to its credit, is moving. The 2026-07-28 spec revision deprecates dynamic client registration in favor of Client ID Metadata Documents, with Cloudflare’s Agents SDK supporting the spec from day zero. A draft proposal would make MCP its own CVE Numbering Authority. The spec repo now carries a Linux Foundation EU Cyber Resilience Act notice with a path for reporting actively exploited bugs. MCP Events was accepted on November 11, adding event-driven capabilities to a request-driven protocol.

The fleet, meanwhile, has barely moved. Only 769 of the 14,664 hosts that report a protocol version, 5.2 percent, speak 2026-07-28, and just 6 of those 769 also sign clients in with CIMD while avoiding its removed or deprecated features. Of OAuth-protected hosts that refused the anonymous probe, 5,366 (63 percent) still rely on dynamic registration with no CIMD support, and 1,292 hosts refuse clients without publishing any discoverable authorization metadata at all.

Even the reference implementations needed patches this month. MCP Bits flags a high-severity advisory where mcp-server-git could execute a command planted in a repository’s own git config via core.fsmonitor during nominally read-only tool calls, plus symlink-swap races in the filesystem server. Five open issues across Codex, Claude Code, Gemini CLI, and Hermes Agent turned out to be one shared bug: the client keeps retrying an OAuth refresh token the server has already rotated. And a Transports WG draft names a structural gap that predates all of it: “MCP services update their tools, prompts, and resources while clients still use earlier definitions… servers cannot reliably distinguish an outdated definition from a bad call, and refreshing does not always resolve the mismatch.”

Two more maturity signals from the same roundup: Isokron found that two agents writing the same object simultaneously lose changes on GitHub’s, Grafana’s, and PostHog’s official servers, and one benchmark ran the same GitHub task three ways and measured the MCP path at roughly 59,000 input tokens against about 1,400 for the CLI path, with a single tool result costing more than all 46 tool definitions combined. None of this is disqualifying. All of it argues for connecting agents to few servers rather than many.

What this means when the agent is buying

Strip the census to its commerce signal and you get three rules.

First, prefer read-only, purpose-built endpoints. An agent that compares products does not need write access, destructive tools, or a 46-tool surface it pays for in context on every call. It needs one narrow answer: is this product what it claims to be?

Second, prefer servers whose text you can audit. Sixty-two percent of the ecosystem lists tools to strangers, and the share of descriptions carrying model-facing instructions is a floor estimate that the scanner authors themselves disavow as complete. A trust endpoint should publish exactly what it returns and nothing that steers the model.

Third, prefer verification over volume. A registry listing proves who published a server. It says nothing about whether the product data inside it survived a review economy where, as we have documented, a single seller-written sentence can move an AI agent’s willingness to pay by double-digit percentages.

That is the design brief GoBuy was built against. The GoBuy MCP server is a read-only trust endpoint: it filters fake reviews, weights authentic ones, and returns a Smart Score from 0 to 100 built on review quality rather than review quantity, plus a shortlist of the top 7 products instead of an endless shelf. Products that hold a score of 80 or above across 90 days earn the GoBuy Verified badge, the same signal available to humans through the GoBuy Chrome extension on Amazon pages. Agents can wire it in minutes via the agent docs.

The census’s quiet conclusion is that the MCP ecosystem’s problem is not malice, it is asymmetry: 59,944 publishers, one anonymous probe away from every agent, and adoption of the security spec at 5.2 percent. Until the fleet catches up, the winning strategy for agent operators is old-fashioned: know exactly which servers your agent talks to, read what they say to it, and make sure at least one of them exists to verify the rest of the internet.

Start at gobuy.ai, or give your agent the trust layer directly at the GoBuy MCP server docs.