Unclear What the added checks refuse is not stated, so it is unclear which existing mcp_tool hooks might stop running.
What
A hook is a command you configure to run automatically at certain points in Claude Code. An mcp_tool hook calls a tool on an MCP server, which is an add-on program that gives Claude Code extra tools. These hooks now face more checks before they are allowed to run:
- A hook may not call a server when a subagent (a helper agent Claude starts for a sub-task) or a background session might have set the hook off.
- In a cloud session, a hook may not call a server that acts on your own computer or browser.
- The checks run both before the call and again after the server reconnects. Before, there were three refusal checks.
- When a hook is refused, it reports messages such as "Hooks can't call MCP server" and "MCP tool hook not run".
- The hook result now includes a
ranOnfield naming the server the call ran on, and the call is registered with a server label.
Why
This narrows which MCP servers hooks may reach, so an existing mcp_tool hook setup that used to work may now be refused. The new ranOn field shows which server actually handled each call.
Names in the bundlemcp_tool
Something disagreesSomething we can check disagrees with this entry, or the writer said they could not settle it.
The writer flagged doubt
What the added checks refuse is not stated, so it is unclear which existing `mcp_tool` hooks might stop running.
Anthropic's release notes agree
Fixed SDK sessions giving an MCP server named by --allowedTools or an mcp_tool hook less than 2s before the first turn when settings lowered…