How does Claude's context window grow within a session?
Tracks token consumption from JSONL usage data — showing peak context size, session size distribution, and total token footprint across all your Claude Code sessions.
npx cc-context
npx cc-context --jsonOr open index.html in a browser and drag in .jsonl files.
- Peak context — highest (input + cache_read + cache_create) seen in a session
- Size tiers — tiny / small / medium / large / huge (150K+ tokens)
- Total tokens — fresh input, cache reads, cache writes, output across all sessions
- Largest sessions — sessions that pushed closest to the 200K context limit
cc-context — How does Claude's context window grow?
Sessions analyzed: 471
Peak context per session
median 90K tokens
p90 167K tokens
max 190K tokens
Session size tiers
tiny 0% (0) <5K — single-turn
small 2% (8) 5K–20K
medium 37% (173) 20K–60K — typical session
large 27% (126) 60K–150K
huge 35% (164) 150K+ — near context limit (200K window)
Total token consumption
input (fresh) 5.1M
cache reads 22.1B (97% of total input)
cache written 668.6M
output 19.4M
grand total 22.8B tokens
Median session uses 90K tokens (45% of the 200K window).
35% of sessions push past 150K tokens — approaching compaction territory.
MIT