Caprock · this session
input cost cut
saved ·
paid of
tokens cut
Where the savings come from
Compression
Prompt cache
Cache hit rate
Output saved
out tokens
Overhead
ms
added latency
Throughput
tok/s
forward p50
Tokens saved by strategy
Cache health
Prefix-freeze busts avoided
Cache busts
Tokens lost to busts
Net cache savings
Cost by client · this session
bar length = share of total input · amber = saved, grey = sent to Bedrock
Client
Compression
In
Out
Saved
$
Cut
Session total
saved ·
tokens ·
Traffic appears here after Claude Code, Cursor, Codex, or another client sends requests through the proxy.
By model
model · % of requests
sent
saved
cut
Tokens saved over time
·
Savings Breakdown
Compression
Prefix Cache Discount
Provider-native caching
Anthropic Subscription Window
5-Hour Window
7-Day Window
Extra Usage (Overage)
Headroom Contribution This Window
Efficiency
Tokens Saved
Compression
Cache Reads
Raw vs Submitted
Forwarded to Anthropic
Saved by Headroom
⚠ Anomalies Detected
OpenAI Codex Rate-Limit Window
Primary
Secondary
Credits
Pay-as-you-go credits balance
GitHub Copilot Quota
∞
Unlimited
Monthly Reset
Month start
Prefix Cache Impact
no activity since restart
Cache Reads (lifetime)
savings not priced yet
pricing needs LiteLLM
Cache Writes
—
no activity since restart
Hit Rate
—
no activity since restart
Cache Busts
—
no activity since restart
Providers
—
with cache data
no activity since restart
Cache Efficiency
Reads (discounted)
Writes
Uncached
Observed TTL Buckets
Provider-reported cache write mix
Bucket Mix
1h
/
5m
Observed write-token split
1h Cache Writes
5m Cache Writes
Compression vs Cache
Tokens saved by compression against cached-prefix tokens its mutations invalidated
Saved by Compression
tokens removed before send
Lost to Cache Busts
Net
saved minus bust losses
Prefix Freeze Net
Cache Miss Attribution
Why turns that expected a prompt-cache hit missed — TTL lapse (consider a longer TTL) vs the cacheable prefix changing
TTL Expiry
Prefix Change
Unknown
stable prefix, within TTL
Total Misses
expected a cache hit, got none
Per-Provider Breakdown
What Caprock Removed
No waste signals detected yet. Data appears after requests are processed.
Savings Over Time
Trend data will appear after multiple requests.
Token Usage
Before Compression
not installed
Proxy Removed
After Compression (sent)
Output Tokens
Providers
No requests yet
Performance
Overhead Range
TTFB Range
Failed Requests
Pipeline Breakdown
Per-Model Token Savings
Exact tokens saved per model
| Model | Requests | Tokens Saved | Tokens Sent | Reduction |
|---|---|---|---|---|
Per-Project Savings
Lifetime totals — attributed requests only
No per-project data yet.
Add to each project's .claude/settings.local.json:
| Project | Requests | Tokens Saved | Saved $ | Savings % |
|---|---|---|---|---|
|
|
Recent Requests
Last 10 — click row to expand
Time
Model
Input
Output
Saved
Latency
Original Tokens
Compressed Tokens
Tokens Removed
Optimization Time
Transforms Applied
Waste Detected
No requests yet. Start using the proxy to see activity here.
Historical Proxy Compression
Durable local savings history from /stats-history
Lifetime Compression Savings
Proxy compression only
Lifetime Tokens Saved
Active Days
Average Saved / Day
Based on persisted daily buckets
Average Saved / Week
Based on persisted weekly buckets
CLI output filtering (lifetime)
Historical Savings Trend
Historical trend data will appear after more saved checkpoints in this granularity.
Actual cost (with Headroom)
Expected cost (without Headroom)
Per-Model Breakdown
Per-model attribution appears for checkpoints recorded after upgrading.
| Model | Tokens saved | Cost with Headroom | Expected cost without Headroom | Saved |
|---|---|---|---|---|
Historical Summary
Latest total
Selected series
Selected points
Average / month
Retention
Daily Savings
Daily rollups will appear after persisted history spans at least one checkpoint.
Weekly Savings
Weekly rollups will appear after persisted history spans multiple days.
Monthly Savings
Monthly rollups will appear after persisted history spans multiple months.
Recent Historical Checkpoints
Cumulative proxy compression savings
No persisted savings history yet
Historical data is written locally after proxy requests save tokens. Keep using Headroom and this view will fill in automatically across restarts.