How the census works
We track six agents: Claude Code, Cursor, GitHub Copilot, OpenAI Codex, Windsurf and Claude Desktop. For each of the nine vendors we read the product pages listed as sources and record whether each agent is named. An agent implied by a general phrase such as coding assistants is not counted. The result feeds our named agent coverage criterion, computed as the count divided by six, times ten, rounded.
Agent by agent
Claude Code is named by seven of nine vendors, the most of any agent. Cursor is named by five, GitHub Copilot by three, OpenAI Codex by four, Windsurf by one and Claude Desktop by one.
The pattern follows how the agents expose control points. Claude Code offers lifecycle hooks and managed settings that vendors can integrate with, and several vendors describe those integrations specifically.
Vendor by vendor
Zenity names four: Claude Code, Cursor, GitHub Copilot and Codex. It also names Cowork, which we do not count as Claude Desktop because the page does not name Claude Desktop.
Bay names three: Claude Code, Codex and Claude Desktop, for which it describes managed settings that lock permission rules, preserve managed hooks, restrict MCP servers and constrain plugin sources.
Noma Security names three: Claude Code, Cursor and Codex, which it discovers through existing EDR or MDM with their MCP servers and skills.
Prompt Security names three: GitHub Copilot, Cursor and Claude Code, which it says it protects from exposing IP or secrets.
Lasso Security names three: Claude Code, through its lifecycle hooks integration, and Cursor and Windsurf, across which its gateway discovers MCP servers.
Koi names three: a Palo Alto Networks blog of 18 August 2026 names Claude Code, Cursor and Codex, and also Antigravity, in describing agent runtime policies.
Harmonic Security names two of the six: its pricing page lists Claude Code and GitHub Copilot, along with ChatGPT Desktop and Microsoft 365 Copilot, among the developer tools and desktop apps its Guide tier covers.
Onyx Security and Bloom Security describe coding assistants, desktop agents or AI tools in general terms and name none of the six on the pages we read.
What the names do and do not prove
A name tells you the vendor wants to be considered for that agent. It does not tell you whether the product only sees the agent, can govern its settings, or decides on each of its tool calls before they run. Those depths differ between products, and between agents in the same product. Bay's own wording, for example, limits in-line enforcement to supported agent surfaces.
The absence of a name proves little either. Onyx Security, which names no agents, ties for the most verdict types documented and scores 8 on stopping an action before it runs. That is why named agent coverage carries 5% of our score and not more.
How to use the census
Start with the agents your people actually run, not the ones vendors list. In the agent risk estimator, tick them and turn on the filter for vendors that name all of them. Then take the shortlist to the three-depth check from lesson 7: for each agent, does the product see it, govern it and decide on its actions?
The list will change
New agents and versions appear every month, and hook interfaces change with them. We will update the census when vendors update their pages. If you see a vendor page that names an agent we have marked as not named, send it through the contact page with the topic Correction.
Agents we do not track yet
Vendors also name agents outside our six, such as Cowork on Zenity's pages, Antigravity on Palo Alto Networks' Koi blog, and ChatGPT Desktop and Microsoft 365 Copilot on Harmonic Security's. We keep the census to six so that every vendor is measured against the same list. If a new agent becomes common across vendor pages, we will add it to the census and recompute the named agent scores for every vendor at once.