Desktop app produces noticeably lower quality responses than CLI for identical prompts

Summary

Running the same prompts in Claude Code Desktop vs CLI yields significantly different quality results, even though both run the same model (Opus 4.6) in local mode.

Concrete example

I asked both environments to update App Store titles and descriptions for my iOS app, using a combination of Fastlane and the App Store Connect MCP server.

Desktop app:

• Failed to use the App Store Connect MCP server correctly
• Complained that the Fastlane setup was incomplete (it wasn't)
• Required constant hand-holding to make progress

CLI:

• Immediately knew how to use both Fastlane and the MCP tools to update everything
• Autonomously discovered it should also investigate and update my landing page
• Proactively researched competitor landing pages for context
• Completed the task end-to-end with minimal intervention

Why this matters

I want to use the desktop app — the GUI, worktree features, and preview buttons are great. But the quality gap makes it impractical for real work right now. The CLI consistently outperforms on identical tasks.

Possible contributing factors

• Desktop loads many more MCP tools simultaneously, bloating the system prompt and diluting model focus
• Context loading (CLAUDE.md, memory, git state) may differ between harnesses
• Tool routing and autonomous exploration seem less capable in the desktop harness