How we label

Every tool gets two labels: what it can do out of the box, and what it can do with our preset. Both are provisional until we've tested them.

Classes

Read-only
Documented as read-only: every tool the vendor lists only reads. Not yet confirmed in our sandbox.
Can write
Can create or change things, but not delete.
Can delete
Can delete, destroy, or run arbitrary commands.
Unknown
We have not found a source for what it can do. Treat it as able to write.
No read-only mode
We found no read-only switch. The credential you give it is the only control.

A server gets the class of its most dangerous tool. A Kubernetes server with 40 read tools and one delete_pod is "Can delete".

Flags

exec
Can run commands (e.g. exec into a pod or run a CLI).
secrets
Can read secrets or sensitive data like logs.
egress
Sends data to a third-party service.
hosted
Vendor-hosted endpoint; you cannot inspect the server code.

Verification status

From docs, not tested yet
We read the project's docs or source and wrote the preset from them. We have not run it. Every preset on the site is in this state today.
Sandbox-tested
Planned. We run the preset against a throwaway cluster/account, call each write tool directly (not just check the tool list), and publish the log with the server version and date.

Why call tools directly? Because in CVE-2026-46519 a read-only mode hid write tools from the list but still let clients call them.

The three layers

  1. Credential: read-only RBAC, IAM role or token scope. The only layer that's a real boundary.
  2. Server flag: the server's own read-only mode. Our presets set this.
  3. Agent rules: Claude Code permissions.deny (mcp__server__tool), Codex default_tools_approval_mode and disabled_tools, Cursor mcpAllowlist in permissions.json. Useful, but Cursor's docs call allowlists "not a security boundary". MCP's readOnlyHint is a hint, not enforcement.

Client formats we follow

Each preset is written in the format the client documents. These are the pages we checked (Nov 3, 2026):

Sources and dates

Every claim links the page we read and the date we read it. GitHub stats come from the GitHub API and are refreshed by a script that flags changes (stars, releases, archived, renamed, license). If a repo hasn't been pushed in 180 days we mark it stale.

What we don't do