Claude was consuming 84.6K tokens firstly of each chat till I made this one change

Do you, like me, obsess over your Claude periods, attempting to optimize every part all the way down to a single token? If so, you’re in good firm — I’ve been attempting every part to stretch my $20 Claude Pro subscription so far as I can. However, all of your tweaks and changes solely work should you’re constructing them on a stable basis. As it seems, the way in which I used to be utilizing Claude was really working in opposition to me and inflating my token utilization from the very first message.

I despatched “Hi,” and the token utilization was already at 85K

Claude’s dealing with numerous context underneath the hood

Claude Cowork used to have a /context command that confirmed what number of tokens a chat had consumed as much as that time. I used it typically to test whether or not I’d reached 200K tokens, at which level I’d summarize every part and move to a new chat to avoid context rot. However, I’d all the time used this characteristic halfway via a chat to see if I used to be nearing my restrict for switching chats. I by no means had the brilliant thought of testing it firstly of a chat, and I solely stumbled upon this not too long ago accidentally.

You see, the /context command appears to have been not too long ago faraway from Cowork. As far as I can inform, it’s now solely out there in Claude Code — each within the CLI and within the desktop app’s Code mode. In truth, should you’re utilizing Code mode within the Claude desktop app, you may entry it immediately from the right-hand panel, so you may all the time regulate your token utilization. I discovered that to be a extremely neat enchancment, so I made a decision to modify from Cowork to Code mode to watch my token utilization extra intently.

So I began a brand new session and despatched “Hi.” Claude replied with one other easy greeting, however to my absolute horror, the panel confirmed that I used to be already utilizing 84.6K tokens. I wasn’t anticipating that in any respect. Sure, I knew Claude had a reasonably large system immediate, so I figured there’d be some baseline overhead — perhaps 5K tokens, or 10K on the very most. But the precise quantity was greater than eight instances what I’d’ve guessed.

Unfortunately, you may’t convey it all the way down to zero

I used to be solely in a position to scale back it by 16K tokens

Seeing 85K/1M tokens consumed after a single “Hi” is startling, however as soon as I appeared on the breakdown, the numbers began to make some sense — even when I didn’t prefer it. The Claude system immediate and built-in system instruments have been consuming up about 30K tokens. Those are baked into the app, and there’s nothing you are able to do about them.

Then there have been Messages, which took up one other 23K tokens. Since the one messages within the chat have been my “Hi” and Claude’s reply of fewer than 10 phrases, that couldn’t account for all of it. My guess is that this consists of reminiscences and context from previous conversations that Claude hundreds firstly of each chat. To check this, I deleted a few of my actually outdated conversations, and that diminished Messages to round 14K tokens. I in all probability may’ve introduced that quantity down additional by disabling the reminiscence characteristic, nevertheless it’s very helpful, so I made a decision to go away it enabled.

The remaining 30K or so tokens got here from Skills and MCP instruments. Now, I’ll admit I’ve put in numerous Skills and Plugins over the months, however certainly they couldn’t be consuming up that many tokens? The thought behind Skills and MCP servers is that solely a short description of every software (and never the software itself) is loaded firstly of a chat, so Claude is aware of what to set off when vital. Surely the descriptions alone couldn’t account for this a lot token utilization?

So I went into Settings and disabled virtually each Connector, Plugin, and Skill. However, that solely diminished token utilization by round 8K. I then went again and deleted every part to see if that made a distinction, nevertheless it seems disabling them was sufficient to take away them from the context. Going scorched-earth didn’t actually assist my trigger.

Overall, in any case that fiddling round, I used to be in a position to convey down the baseline token utilization from 84.6K to 68.7K — a discount of roughly 16K tokens. Sure, it’s higher than nothing and would in all probability assist prolong a session by one or two further messages, however I definitely would’ve appreciated to get that quantity down to simply 30K, accounting just for the system prompts and instruments.

Ultimately, I needed to swap apps to resolve the issue

This is why energy customers love Claude Code CLI

Both Claude Code and Cowork mode within the desktop app technically run the Claude Code CLI underneath the hood — in any case, it was Anthropic’s first harness of this sort. As such, I wished to return to the supply and check whether or not the identical subject continued.

So I opened a Claude Code session within the terminal, logged in with the identical Claude account, despatched “Hi,” and ran the identical /context command. This time, I noticed a comparatively affordable token utilization of 46.7K, even earlier than making any tweaks or changes. The token distribution was additionally fairly completely different. The system instruments price roughly the identical in each apps, however the CLI’s system immediate accounted for simply 3K tokens, in comparison with virtually 10K within the desktop app.

The MCP instruments barely contributed to context utilization. For comparability, the desktop app confirmed 13.9K tokens attributed to MCP instruments. Part of that got here from the 2 connectors I stored enabled — Claude in Chrome and Filesystem — whereas the remaining got here from no matter extra tooling the desktop app consists of. However, the truth that MCP instruments accounted for lower than 1K tokens within the Claude Code CLI means that it doesn’t want a lot extra tooling to do its job.

This means the desktop app isn’t simply heavier on system sources due to its graphical layer. It’s additionally heavier by way of token utilization as a result of it ships with extra tooling and a bigger system immediate.

That stated, the actually cool factor concerning the CLI is that it allows you to configure your setup on a per-folder foundation. It shops settings domestically in every folder, together with a particular assortment of Skills and MCP servers. To check this, I stripped a second folder of every part and launched Claude Code there. The similar “Hi” introduced the token depend down to simply 36.9K, whereas my different folder retained its full setup. This means you may doubtlessly have one folder optimized for programming with Claude Code and one other for advertising and marketing work. You don’t must allow (or disable) each software throughout the board and find yourself with a generic, suboptimal workflow.

Anthropic wants to enhance their desktop app

I used Claude Code via the CLI for months earlier than Anthropic launched Cowork mode within the desktop app. It allowed you to work with your system files simply as you can with Claude Code, however via a graphical interface. At that time, I switched as a result of I discovered the graphical app extra aesthetically pleasing — the CLI makes every part really feel very claustrophobic. However, my current testing revealed that the CLI nonetheless has some main benefits over the graphical app, particularly on the subject of modularity and person management.

As such, I’d prefer to ask Anthropic to convey the identical degree of flexibility we’ve with the CLI app to the desktop app. And if that’s an excessive amount of hassle, a minimum of make the CLI extra nice to make use of, with higher spacing and margins.