DEV Community

Yurukusa
Yurukusa

Posted on

I Had No Idea How Full My Claude Code Context Window Was. So I Built a Tool.

#claudecode #ai #productivity #cli

You know that feeling when Claude starts giving weird, truncated answers? Or when it seems to "forget" earlier context in a long session?

That's your context window filling up.

Claude Sonnet has a 200,000 token context window. When it's full, older messages get dropped. Your carefully set-up context — the files you loaded, the constraints you established, the codebase you explained — silently disappears.

I've been running Claude Code for 60+ days in an autonomous loop. Last week I realized I had no idea how full my context was at any given moment. So I built cc-context-check.

What it does 

npx cc-context-check
Enter fullscreen mode Exit fullscreen mode 
cc-context-check — Context window usage across sessions

Context limit: 200.0k tokens (Claude Sonnet/Opus)

🟢 ~/projects/my-app    [a3f9c12] just now · 12.4 MB
   ████████████░░░░░░░░░░░░░░░░░░ 40.1% used
   80.2k input · 1.2k output · 119.8k remaining

🟡 ~/                   [b7d44e1] 2h ago · 5.9 MB
   █████████████████████░░░░░░░░░ 71.5% used
   143.0k input · 89 output · 57.0k remaining
   △ Warning: Context is getting full — consider /compact
Enter fullscreen mode Exit fullscreen mode

Color coded: 🟢 under 70% → 🟡 70-85% (consider compacting) → 🔴 85%+ (compact now)

How it works

Claude Code stores session transcripts as .jsonl files in ~/.claude/projects/. Each line is a message. Each assistant message includes a usage object:

{
  "usage": {
    "input_tokens": 1,
    "cache_read_input_tokens": 79927,
    "cache_creation_input_tokens": 917,
    "output_tokens": 158
  }
}
Enter fullscreen mode Exit fullscreen mode

The real context usage is input_tokens + cache_read_input_tokens + cache_creation_input_tokens. The input_tokens: 1 is misleading — the cache tokens are where your actual context lives.

cc-context-check reads the last 64KB of each session file, finds the most recent usage data, and calculates your true fill percentage.

When to use /compact

The tool shows all your active Claude Code sessions across all projects. So you can see at a glance:

  • 40%: You're fine. Keep going.
  • 71%: Getting full. /compact will help if you're starting a new task.
  • 87%: Compact soon. You're losing context.

Options 

npx cc-context-check --all    # Show top 20 sessions instead of 5
npx cc-context-check --json   # JSON output for scripting
Enter fullscreen mode Exit fullscreen mode

Part of cc-toolkit

This is tool #43 in cc-toolkit — 43 free tools for Claude Code users. Zero dependencies, runs in browser or via npx.

Source: github.com/yurukusa/cc-context-check

More tools: Dev Toolkit — 56 free browser-based tools for developers. JSON, regex, colors, CSS, SQL, and more. All single HTML files, no signup.

Top comments (1)

Collapse
 
nyrok profile image
Hamza KONTE •

Useful tool — context blindness is one of the most common silent killers of long Claude Code sessions. You don't notice the window filling up until the output quality drops.

Two complementary approaches: your tool for visibility on the context side, and structured prompts on the input side. Compact, well-structured prompts (explicit role, objective, constraints in XML rather than prose) use context tokens more efficiently and leave more room for actual work.

I built flompt (flompt.dev) for the latter — visual prompt builder that compiles to Claude-optimized XML, also an MCP server: claude mcp add flompt https://flompt.dev/mcp/ Free and open-source.