ContextMesh

ContextMesh — Intelligent Context Layer for Claude Code

One coding session. Infinite memory. 90% fewer tokens.

ContextMesh is a **pure transparent proxy + memory engine** that sits between Claude Code and the Anthropic API. It automatically compresses token waste, maintains a persistent knowledge graph of your work, and injects Aider-style codebase repomaps — all without changing how you use Claude.

**No MCP. No wrappers. Just run `contextmesh init` once globally. Then use `claude` normally forever.**

---

How it works (Pure Proxy Architecture)

Everything happens automatically in the proxy. Claude doesn't need to learn any new tools or call any MCP endpoints.

Your Terminal (just run `claude`)
    │
    ▼
┌──────────────────────────────────────────┐
│         ContextMesh Smart Proxy          │  ← Intercepts every request globally
│  ┌─────────────────────────────────┐     │
│  │  Aider-Style RepoMap Injector   │     │  ← Injects AST map into system prompt
│  │  Session Resumption Injector    │     │  ← Injects last session memory
│  │  RTK Output Compressor          │     │  ← Crushes terminal noise
│  │  Anti-Context Auto-Flusher      │     │  ← Removes old resolved turns
│  └─────────────────────────────────┘     │
│             ↓ Stats Writer ↓             │
└─────────────┼──────────────┼─────────────┘
              │              │
      To Anthropic API     To Local SQLite (for dashboard & stats)

---

Features

Feature	What it does	Savings
RTK Output Compressor	Intercepts massive grep, npm test, cat outputs and crushes the middle noise	Up to 90% on tool outputs
AST Repo-Map Injector	Parses your entire codebase with Tree-sitter and silently injects a dense structural map into Claude's system prompt on Turn 1	95% on code reads
Anti-Context Auto-Flusher	Silently drops old resolved tool calls from history when context bloats past 150k chars	30–60% on long sessions
**Session R