Context Brain
See what Claude Code is holding in its head, and what to cut. Context Brain turns a Claude Code session into a live 3D brain: every file it read, every shell output, rule, skill and tool result becomes a neuron in its own region, and you watch light flow from the context into each answer. A built-in context doctor then tells you where tokens leak and how to get them back.
▶ Live demo (no install, replays a sample session) · 한국어
npx github:workdd/context-brain
That's it. It opens your browser and follows your most recent Claude Code session in real time, staying within that project. Nothing to configure: it only reads the session logs Claude Code already writes.
What it answers
| Question | Where you see it |
|---|
| What is filling my context? | Region badges with token totals, neuron size, name tags on the biggest items |
| What did this answer actually use? | Lines from each item used in the turn to the Response region (thickness = tokens returned, brighter = the answer names it). A line is a record of use, not live attention |
| What is just taking up space? | Red pulsing neurons: 800+ tokens, not used again for more than 2 turns |
| How close am I to auto-compaction? | The ring in the corner, against where this model auto-compacted before. If past auto-compactions are too spread out to call a threshold, it shows the range instead of a percentage |
| When and why did the context grow? | The context history strip under the brain: size per API call by region, the biggest jumps labeled with their cause, compactions marked |
| What should I change? | Context doctor: each leak with its size and a concrete fix, in the browser, the terminal and the Claude Code status line |
Context doctor
| Finding | Fix it suggests |
|---|
| Same file read in full more than once | Read with offset/limit, Grep first, summarize once |
| Oversized tool output (4k+ tokens in one result) | Pipe through head/grep/tail, quiet flags |
| Items never used again (800+ tokens, idle 2+ turns) | /compact or /clear at the task boundary |
| MCP server instructions loaded but never called | Disable the server for this project |
| Tool schemas loaded but unused | Name the tool in the prompt |
| Repeated tool failures | Tell Claude the cause |
| Large fixed context (rules, skills, hooks) | Trim CLAUDE.md, disable unused skills and plugins |
| About to auto-compact | /compact now with a note on what to keep |
The "reclaimable now" number counts only what leaves the context if you act: unused items still in context, unused MCP instructions and unused schemas.
$ npx github:workdd/context-brain doctor
Context doctor · Fix login 401
context 232k / compacts at 810k (29%) · 24 turns
reclaimable now: ~15.6k tokens
HIGH 8 items never used again · 13.1k
Run /compact or /clear at the task boundary to drop this. ...
- src/orders/orders.service.ts 3.2k · idle 2 turns
Project doctor
One session can be a coincidence. The project doctor reads the project's 20 most recent sessions and keeps only what repeats:
| Finding | Why it is a strong signal |
|---|
| MCP server loaded in many sessions, never called | Its instructions are paid for in every session for nothing. Safe to disable for this project |
| File re-read in session after session | A few lines in CLAUDE.md about it save reading it whole each time |
| Command with consistently large output | Grouped by program (npm, gh, pytest …). File readers like cat and sed are excluded |
| Fixed context in every session | What is spent before your first prompt: system prompt, skill list, CLAUDE.md files, hooks |
Click "Analyze recent sessions" under the brain, or run doctor --project in the terminal. Click a session bar to open that session.
In the Claude Code status line
Add this to ~/.claude/settings.json and the reclaimable amount follows you in every session:
{ "statusLine": { "type": "command", "command": "npx -y github:workdd/context-brain statusline" } }
🧠 ctx 232k (29%) · reclaim 16k · 2 high
It reads transcript_path from the status line input and caches the compaction point for 10 minutes. For speed, clone the repo and point command at node /path/to/context-brain/bin/context-brain.js statusline.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
