Hey,
For the last few days, I have had a strange issue.
Claude Code injects a massive token amount to Context under Messages, even when I just start a new session and say "Hey!" - and massive means even 100k.
For example, I turned off all my plugins, MCPs, and skills, started a fresh session on a small repository.
- First thing, verify the context:
9.8k/200k tokens (5%), where⛁ Messages: 8 tokens (0.0%)- looks good, - Wrote
Hello!, so Claude will reply with a 2-3 sentence, - Check context again:
68k/200k tokens (34%), where⛁ Messages: 58.2k tokens (29.1%)
I'm wondering what he is injecting there, some dead leftovers? With this now, using Claude literally is impossible as it fails after a couple of commands with overloaded context. Has anyone had something similar before?
Talking with Claude itself does not help at all, unfortunately. Verbose mode gives me only info about the number, but nothing more.
The 101k tokens in "Messages" is almost entirely from this conversation itself — specifically the two /context command outputs and my responses reproducing them.
Here's what's in there:
1. First /context invocation — the raw ANSI-colored output from the command (all MCP tools listed) plus my response where I reproduced the entire thing as a markdown table, doubling the token cost.
2. "Hello!" exchange — trivial, ~50 tokens.
3. Second /context invocation — same raw output again, plus my response reproducing it again as a full markdown table.
The culprit is the MCP tools list. That section alone has 49 tools, each with a name, server name, and token count. Listing them twice in raw command output + twice in my formatted responses adds up fast.
What inflated it further: my responses converted the already-large command output into markdown tables, which are token-heavier than plain text due to | delimiters on every cell.
Practical takeaway: /context is expensive to run repeatedly in a session because each invocation + response adds ~20-30k tokens to the conversation history. Use it sparingly, or use /clear after checking context if you
don't need the output in history.
I did the fresh reinstall - started Claude in a small repository with nothing configured except my ENVs.
❯ How to lock the screen on a MacBook?
:black_circle_for_record: Press Command + Control + Q to instantly lock your screen.
Alternatively: Apple menu () → Lock Screen.
✻ Churned for 2s
❯ /context
⎿ Context Usage
⛁ ⛁ ⛁ ⛀ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ claude-sonnet-4-6
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ 50.8k/200k tokens (25%)
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛶ ⛶ ⛶ ⛶
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ Estimated usage by category
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ System prompt: 5.7k tokens (2.8%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ Skills: 618 tokens (0.3%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ Messages: 44.6k tokens (22.3%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ Free space: 116.2k (58.1%)
⛶ ⛶ ⛶ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ Autocompact buffer: 33k tokens (16.5%)
⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝
Skills · /skills
Better, but still not good tbh. 44.6k tokens in Messages, when I have a fresh install, small repo, no `CLAUDE.md` files, no skills, no MCPs. That test was on the latest stable from brew:
❯ claude --version
2.1.123 (Claude Code)
Upgrading to the latest (2.1.133) did not help. Downgrading to 2.1.119 also did not help, but downgrading to 2.1.92 did change something. Now, when saying simply hello world to Claude, I got a reply, and it ate 50k tokens, but the detailed summary looks strange:
❯ /context
⎿ Context Usage
⛁ ⛁ ⛁ ⛀ ⛀ ⛶ ⛶ ⛶ ⛶ ⛶ claude-sonnet-4-6
50.2k/200k tokens (25%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶
Estimated usage by category
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ System prompt: 5.9k tokens (2.9%)
⛁ Skills: 349 tokens (0.2%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ Messages: 294 tokens (0.1%)
⛶ Free space: 160.5k (80.2%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛝ Autocompact buffer: 33k tokens (16.5%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶
⛶ ⛶ ⛶ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝
⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝
Skills · /skills
We see here 50.2k tokens used, but when we sum up the rest of the groups, we'll have less than 7k. I know it's "estimated" but still, the difference is huge.
Has anyone had something similar? I do not have any other ideas how to fight with that right now, but this literally makes my Claude unusable - after two or three prompts, it fails with an out of Context issue...
Best,
D.