Skip to content
IsoKron

Deleting tools are not marked non-destructive

Tools that delete, overwrite, cancel, revoke or reset must set destructiveHint: true; with false, clients run them without a confirmation prompt, and both directories treat that as incorrect annotations.

Check annot-delete-marked-safe · Annotations (both directories) · default severity: blocking · rubric 2.1.0

What fails

The name reads as deleting, overwriting, cancelling, revoking or resetting, but the tool sets destructiveHint: false, so Claude would not prompt before calling it. Both platforms ask for true here; OpenAI adds that being able to undo the action does not justify false. The name match is a heuristic.

The rule

For writes, use true for potentially destructive or irreversible effects, such as deletion, overwriting, cancellation, access revocation, or irreversible sends or transactions.

Source: OpenAI plugin guidelines, section 6.3

How to fix it

Set "destructiveHint": true.

  1. Set destructiveHint: true on every delete, remove, purge, overwrite, cancel, revoke and reset tool.
  2. Do not set false because the action can be undone: OpenAI says reversibility does not justify it.
  3. Consider a dry_run argument as well (see the delete-guard check).
Before
{ "name": "delete_issue", "annotations": { "readOnlyHint": false, "destructiveHint": false } }
After
{ "name": "delete_issue", "annotations": { "readOnlyHint": false, "destructiveHint": true } }

TypeScript SDK

TypeScript
import { McpServer } from '@modelcontextprotocol/server';
import * as z from 'zod';

const server = new McpServer({ name: 'issues', version: '1.0.0' });

server.registerTool(
  'delete_issue',
  {
    title: 'Delete issue',
    description: 'Delete one issue permanently. Use only when the user explicitly asks to delete an issue.',
    inputSchema: z.strictObject({
      issue_id: z.string().describe('Issue ID, for example ISS-42'),
    }),
    annotations: { readOnlyHint: false, destructiveHint: true, openWorldHint: false },
  },
  async ({ issue_id }) => ({ content: [{ type: 'text', text: 'Done' }] }),
);

Python SDK

Python
from typing import Annotated

from pydantic import Field
from fastmcp import FastMCP
from mcp.types import ToolAnnotations

server = FastMCP("issues")

@server.tool(
    title="Delete issue",
    annotations=ToolAnnotations(readOnlyHint=False, destructiveHint=True, openWorldHint=False),
)
def delete_issue(
    issue_id: Annotated[str, Field(description="Issue ID, for example ISS-42")],
) -> str:
    """Delete one issue permanently. Use only when the user explicitly asks to delete an issue."""
    return "Done"

Run against @modelcontextprotocol/server 2.4.0, mcp 2.3.0 and fastmcp 4.2.2 on 2026-11-18; the checker confirms this check passes on the tool list each snippet produces.

Questions

Our delete goes to a trash bin. Is it still destructive?

Yes. OpenAI’s guidelines say the ability to undo does not justify destructiveHint: false for deletion, overwriting or cancellation.

How does the checker know a tool deletes?

From its name (delete, remove, purge, overwrite, cancel, revoke, reset and similar). It is a heuristic; a false match can be ignored.

Further reading

How this check is graded

It counts toward: Anthropic connector directory, OpenAI plugin directory. See the rubric.

Last verified against OpenAI plugin guidelines, section 6.3 on 2026-11-18. The rule above is quoted in the source’s own words. This page as Markdown.

Check your tool list: paste your tools/list, nothing is sent anywhere