Context Window State
200,000 tokens max
MCP Tools
0 tk
System
0 tk
User
0 tk
Assistant
0 tk
Tools/RAG
0 tk
MCP
System
User
Assistant
Tool
RAG
Token Blocks each ▪ = ~50 tokens · ■ = MCP tool defs
◀ MCP RESERVED
Message Stream
⚙ Processing Pipeline
📥
Input Tokenisation
Text → BPE tokens
📐
Context Assembly
Stack all turns in order
🔢
Positional Encoding
Position per token
🧠
Attention Pass
All tokens attend to all
✨
Token Generation
Autoregressive decode
📤
Output + KV Cache
Append & cache keys
📊 Live Stats
Turns
0
messages
Tokens/s
0
generation
Fill
0%
of window
MCP Overhead
0 tk
pre-loaded
🧠 Context Quality
Effective capacity breakdown
MCP tools
0%
System
0%
Conversation
0%
Free space
100%
✓ Context healthy — ample space for conversation
🔥 Attention Heatmap
💰 Live Cost Counter
Cache: OFF
$0.0000
USD this session
Input tokens
0
$0.0000
Output tokens
0
$0.0000
Cache reads
0
-$0.0000
claude-opus-4-6 · $15/$75 per MTok
⚡ KV Cache State
keys+values stored per layer
Hit rate
0%
Cached
0K
KV Entries (each cell = ~100 tokens)
Cached (reused)
New (computed)
Evicted
KV cache stores computed key/value pairs so previous tokens aren't re-processed each generation step.