[BUG] "Tool access" setting has no effect — MCP tools always load eagerly in Claude Code Desktop

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

The "Tool access" setting (Settings → Features → Tool access) has no effect on how MCP tools are loaded in Claude Desktop.

Two problems:

1. Toggling the setting changes nothing. Switching between "Load tools when needed" and "Tools already loaded", then starting a new conversation, produces identical /context output. MCP

tool schemas are always loaded eagerly regardless of the selected option.

1. Desktop differs from CLI. Claude Code CLI correctly defers MCP tool schemas (loaded on-demand via ToolSearch, 0 tokens at session start). Desktop always loads them eagerly — ~75k tokens

for 155 tools in my setup.

This causes Desktop sessions to have 4× higher baselines than CLI (~92k vs ~22k). For agentic tasks with many tool iterations, the overhead compounds across every turn via the prompt cache.

What Should Happen?

1. Toggling between "Load tools when needed" and "Tools already loaded" should produce different /context output in new conversations.

1. When "Load tools when needed" is selected, Desktop should defer MCP tool schemas the same way CLI does — the /context output should show "MCP tools · /mcp (loaded on-demand)" with ~0

tokens, matching CLI behavior.

Error Messages/Logs

Steps to Reproduce

Setup: macOS, Claude Desktop, multiple MCP connectors enabled (e.g. Linear, Datadog, Slack).

Test 1 — Setting toggle has no effect:

1. Go to Settings → Capabilities → Tool access → select "Load tools when needed"
2. Start a new conversation, type "say hi", wait for response
3. Run /context — note the "MCP tools" line
4. Go to Settings → Capabilities → Tool access → switch to "Tools already loaded"
5. Restart Claude Desktop
6. Start a NEW conversation, type "say hi", wait for response
7. Run /context — "MCP tools" line is identical to step 3

<img width="678" height="994" alt="Load tools when needed" src="https://github.com/user-attachments/assets/b4b90a66-a973-4085-b0b0-572be691277f" />

<img width="662" height="988" alt="Tools already loaded" src="https://github.com/user-attachments/assets/f4defca0-ee8f-48d7-8619-70db6089b380" />

Test 2 — Desktop vs CLI comparison:

1. Open Claude Code CLI in the same directory: claude
2. Type "say hi", wait for response
3. Run /context — observe "MCP tools · /mcp (loaded on-demand)" with 0 tokens

<img width="717" height="474" alt="CLI mcp tools" src="https://github.com/user-attachments/assets/0a9b23d3-275f-4a88-a9be-0d8f459a564b" />

Results:

• Setting A ("Load tools when needed"): MCP tools = 33.6k
• Setting B ("Tools already loaded"): MCP tools = 33.6k — identical
• CLI (same machine, same connectors): MCP tools = 0 (on-demand)

Claude Model

Not sure / Multiple models

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.109 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Other

Additional Information

_No response_