Claude Code /context Guide: Find and Reduce Desktop App Token Overhead

Treat those numbers as one person's observations. They are not an Anthropic benchmark, and nobody has reproduced them under controlled conditions. The comparison is still useful for anyone who runs Claude Code on a Windows, macOS or Linux workstation. It also shows how to find out where your own overhead comes from.

What the author reported​

The author is Dibakar Ghosh. He started a fresh session in the desktop app's Code mode and sent a one-word greeting. Claude replied with a short greeting of its own. The context panel then showed 84.6K tokens used against a 1M window.

He had expected 5K to 10K at most. His rough breakdown of the desktop reading:

• System prompt and built-in tools: about 30K tokens. He says this is baked in and can't be changed.
• Messages: about 23K tokens. The chat held only two short messages, so this figure is odd. He guesses it includes memory and context from past conversations. After deleting some very old conversations it fell to about 14K. He kept the memory feature switched on.
• Skills and MCP tools: about 30K tokens.

He then disabled almost every connector, plugin and skill, which saved roughly 8K tokens. Deleting them afterwards changed nothing more, so disabling was enough to take them out of context. His baseline fell from 84.6K to 68.7K, a cut of about 16K. He says that is still far from the roughly 30K he would see if only the system prompt and tools remained.

The terminal comparison​

He then ran Claude Code in a terminal, signed in with the same account, sent the same greeting and ran /context. The total was 46.7K before any tweaks. His breakdown of the difference:

• System tools: roughly the same cost in both places.
• System prompt: about 3K tokens in the CLI, against nearly 10K in the desktop app.
• MCP tools: under 1K in the CLI, against 13.9K in the desktop app. Two connectors he left enabled, Claude in Chrome and Filesystem, account for part of that. The rest he attributes to extra tooling the desktop app ships with.

He also tried a second folder that he had stripped of its skills and MCP servers. Claude Code launched there started at 36.9K, while his original folder kept its full setup. He concludes that you can keep one lean, project-specific setup for coding and another for, say, marketing work. Anthropic's CLI reference does list a --safe-mode flag that starts with customizations disabled, including skills, plugins and MCP servers. That makes a similar before-and-after comparison easy to run.

Caveats​

The article doesn't give app or CLI version numbers, the model used, full settings, or repeated trials. Several of its claims are the author's inference:

• Desktop Code mode's larger prompt and extra tools cause the higher figure.
• Old conversations are why the "Messages" category was inflated.
• The folder alone explains the drop to 36.9K.

The author also says /context was recently removed from Cowork. I couldn't confirm that in the documentation I could reach.

Other published guidance shows that large startup numbers aren't unusual. Anthropic's docs say running /context gives a live breakdown by category with optimization suggestions, including which CLAUDE.md and auto memory files loaded. One independent tip site shows a 1M-window session with a 21k system tools line and a 4.4k system prompt. Another guide notes that the usable space is always smaller than the headline number, because the system prompt, tools, memory, and your CLAUDE.md occupy part of the window before you type anything. Those examples come from different setups, so they can't be compared directly with the author's.

Context tokens and your plan limit​

It's tempting to read 85K tokens as 85K tokens of your Pro allowance gone. The two measures aren't the same thing. One explainer describes the context window as per-session working memory, while the usage limit is a plan-level cap over time. Anthropic's help center says usage depends on several factors, including message length, conversation length, tool usage, model choice and effort level. It also says its system includes caching. A smaller startup context may stretch a session, as the author suspects, but the context panel alone doesn't prove how much allowance you'll save.

How to check your own setup​

1. Open Claude Code in the app's Code mode or in a terminal. On Windows, Anthropic's docs describe running it natively or inside WSL.
2. Start a fresh session and send a trivial message.
3. Run /context and note the categories before changing anything. Anthropic's docs describe it as a live breakdown by category.
4. Change one thing at a time. Disable connectors or skills you don't need for the task, then run /context again.
5. Compare surfaces if you can. Run the same test in the desktop app and in the terminal with the same account.
6. Record the app or CLI version so you can tell whether a later change comes from an update.

Anthropic's support pages draw a useful line between remote connectors and local desktop extensions. Remote connectors are cloud services that work everywhere you use Claude. Local extensions run on your machine and are only available in Claude Desktop and Claude Code. Neither page says how many tokens each costs, so measure that yourself.

I would not copy the author's habit of deleting old conversations to trim "Messages". It rests on his own explicit guess about what that category contains. The memory feature is also something he chose to keep, so weigh it against its usefulness.

What this means for developers​

The practical lesson is to check before you optimize. Fixed overhead from the system prompt and built-in tools can't be removed. Skills, MCP servers and memory can be, and they are where the author found savings. If you use Claude Code on a Windows development machine, a five-minute /context check will show whether your own setup has the same imbalance. The author's request to Anthropic is for the desktop app to offer the per-folder flexibility the CLI has. Whether that happens is up to Anthropic.