[FEATURE] Claude Desktop: opt-in setting to auto-approve write/modifying MCP tools per server

[petrklucka]

Preflight Checklist

• I have searched existing requests and this feature hasn't been requested yet
• This is a single feature request (not multiple features)

Problem Statement

For write/modifying MCP tools (no readOnlyHint), the "Allow for all tasks"
button is disabled, so every single call must be confirmed manually. When
driving a local MCP server that performs many mutating operations in a row
(e.g. an Enterprise Architect / modeling server creating dozens of elements
and connectors), this means hundreds of manual confirmations per session,
which makes batch/automation workflows impractical in Claude Desktop.

Proposed Solution

Add an explicit, opt-in setting in Claude Desktop (default OFF), e.g.
"Allow modifying (write) MCP tools without per-call approval".

Today write/modifying MCP tools (no readOnlyHint) cannot be granted
"Allow for all tasks" — it's greyed out — so every mutating call must be
confirmed manually. Read-only tools can be auto-approved; write tools cannot,
and there is no global or per-server setting to change this.

The setting should keep it safe via:

1. Off by default; user must consciously enable it.
2. Granularity per MCP server (ideally per tool), not one global switch.
3. Scope option: restrict to local (stdio) servers only, exclude remote connectors.
4. Clear warning at enable time (same wording as the per-call dialog).
5. Session-/time-scoped option (auto-revert) + one-click revoke.
6. Audit/log of auto-approved write calls.

This brings the Desktop UI to parity with Claude Code (CLI), which already
supports a permission allowlist (mcp__server__*).

Alternative Solutions

No response

Priority

Critical - Blocking my work

Feature Category

CLI commands and flags

Use Case Example

I run a local stdio MCP server for Enterprise Architect modeling. A single
modeling session creates dozens of elements and connectors in a row
(element_create, connector_create, …). Each is a write tool, so Claude Desktop
prompts for approval on every call — hundreds of manual confirmations per
session — which makes batch/automation workflows impractical.

With an opt-in "allow write MCP tools for this server" setting, I could enable
it consciously for my trusted local EA server and run the batch uninterrupted,
exactly as I already can in Claude Code via the mcp__server__* allowlist.

Additional Context

Complements #25966 (auto-approval of read-only MCP tools); this request covers
the write-tool case with explicit opt-in safeguards.

Environment: Claude Desktop on Windows 11, local stdio MCP server.
Current workaround: use Claude Code CLI with an mcp__server__* allowlist,
which has no equivalent in the Desktop UI.
Rejected alternative: annotating write tools as readOnlyHint:true — semantically
wrong and unsafe.