Your Claude session is doing real work — reading source files, grepping for symbols, fetching docs, spawning agents. And every single one of those tool calls is dumping raw data straight into the context window.
By the time you're 60% through a complex task, the window is stuffed with file contents that already answered their question, grep output nobody needs anymore, and a web page from two subtasks ago. Claude starts forgetting the beginning. You lose the thread. The session dies before the work is done.
That's context rot. Every developer using Claude Code hits it. ctx-tree aims to eliminate it.
Every tool call — native and MCP — is intercepted by a hook, the result is stored in a persistent SQLite property graph, and Claude gets back a compact nodeId reference instead of raw bytes. The same file read next week costs zero tokens. It's already in the graph.
Before / After
| ❌ Without ctx-tree | ✅ With ctx-tree |
|---|
Read auth.ts → 4,200 tokens, session only
gone when context scrolls
re-read it again next session
Grep 'handleToken' → 8,400 tokens, session only
8 files inline, all forgotten
same grep, same cost, next week
WebFetch API docs → 11,000 tokens, session only
full HTML, 90% noise
re-fetched every session
Session stalls at ~40k tokens.
Start over.
|
Read auth.ts → 60 tokens (nodeId)
stored in graph, forever
free recall next session
Grep 'handleToken' → 160 tokens (8 nodes)
searchable across sessions
neighbors found in one call
WebFetch API docs → 110 tokens (reference)
compact a11y tree stored
never re-fetched
Session runs to completion.
Keep going.
|
| Operation | Without | With | Saved |
|---|
| Read file | ~4,200 tok | ~60 tok | 98.6% |
| Grep (8 files) | ~8,400 tok | ~160 tok | 98.1% |
| WebFetch | ~11,000 tok | ~110 tok | 99.0% |
| MCP tool | ~3,600 tok | ~70 tok | 98.1% |
Install
# Add the marketplace — one-time global setup
/plugin marketplace add joeblackwaslike/agent-marketplace
# Install ctx-tree — MCP server, all 9 hooks, skills configured automatically
/plugin install ctx-tree@agent-marketplace
That's it. Start a new Claude Code session. Every tool call is already being intercepted.
Requirements
macOS arm64 or Linux x64/arm64 · Bun ≥1.1 · rg ≥13
brew install bun ripgrep # macOS
curl -fsSL https://bun.sh/install | bash && apt install ripgrep # Linux
What you get
98–99% context reduction per operation. Every read, search, and fetch that used to cost thousands of tokens now costs tens. Operations you've already run cost nothing — the graph remembers.
Zero configuration. Total interception. Ten hooks fire automatically across every native Claude Code tool and every MCP tool you have installed. Nothing slips through. No per-project setup. No flags to flip.
Knowledge that outlasts the session. Prior reads, searches, and web fetches are available next week via ctx_tree_search and ctx_tree_compose. The graph grows with every session. The context stays lean.
How it works
