Search-First 3K Ladder
Eliminates LLM context bloat. Instead of loading 735 tool schemas (~50K tokens), a ~3K token discovery core uses BM25 + ONNX micro-embeddings + RRF to activate tools lazily.
735 tools across 36 domains for AI agents: unifying network packets, DOM, V8/WASM, native memory, headless Unicorn JNI emulation, and kernel syscalls with a 3K-token profile ladder.
Add this single configuration snippet to your MCP client (Claude Desktop, Cursor, Claude Code, Windsurf, or Antigravity):
{
"mcpServers": {
"jshook": {
"command": "npx",
"args": ["-y", "@jshookmcp/jshook"]
}
}
}Production Profile Tiering
By default, the server runs under MCP_TOOL_PROFILE=search, consuming only ~3,000 tokens of initial context. The agent automatically discovers and activates domains and specific tools on demand via search_tools / route_tool / activate_tools.
Modern security challenges (Web SPAs, Electron desktop apps, hybrid Android apps) distribute critical logic across network protocols, V8 JIT optimizations, WebAssembly runtimes, native C/C++ .so dynamic libraries, and kernel syscalls.
Single-purpose tools leave AI agents blind across system boundaries. JSHookMCP provides a unified, cross-domain causal evidence graph:
[Network: HTTP/2 / gRPC / Protobuf / TLS Keylog]
↓
[DOM / UI: Camoufox CDP / Canvas / WebGL Interception]
↓
[JS Runtime: AST Deflat / V8 Turbofan JIT / Heap Diffing]
↓
[WASM Virtual Machine: Wasm2C Decompilation / VMP Tracing]
↓
[Internal IPC: Chromium Mojo IPC Sniffing & Mocking]
↓
[Native Memory: AOB Signature Scan / Multi-level Pointer Maps]
↓
[Headless JNI Emulation: Unicorn ARM64 / Automatic JNI Environment Mocking]
↓
[OS Kernel: POSIX Syscalls / Mach Traps / Win32 VEH / eBPF Tracing]| Feature | Browser MCPs (Playwright/Puppeteer) | Binary MCPs (Ghidra/IDA) | Traditional Dynamic Tools (Frida/CheatEngine) | JSHookMCP |
|---|---|---|---|---|
| Tool Count & Scope | ~10-20 tools | ~10-30 tools | Standalone scripts | 735 tools across 36 domains |
| LLM Context Footprint | Few thousand tokens | Tens of thousands (bloat) | No native agent interface | Search-First 3K Ladder (Dynamic Activation) |
| Cross-Domain Causality | ❌ DOM / Network only | ❌ Single binary only | ❌ Manual human synthesis | ✅ Automatic causal correlation & evidence export |
| Headless Android JNI | ❌ None | ❌ None | ❌ Requires device/emulator | ✅ In-process Unicorn ARM64 emulation |
| Flutter / Dart AOT | ❌ None | ❌ Poor | ❌ Very difficult | ✅ Snapshot header dump, SMI scan, call graphs |
| Chromium Mojo IPC | ❌ None | ❌ None | ❌ None | ✅ Full message interception & mocking |
| AOB Scan & Pointer Maps | ❌ None | ❌ Static only | ⚠️ GUI client only | ✅ Fully exposed to AI agents |
| Domain Category | Domains Included | Key Capabilities |
|---|---|---|
| Causal Correlation | cross-domain, core, trace, session | Evidence graphs, execution trace recording, dynamic tool routing |
| Headless Emulation | native-emulator (nemu), dart-inspector, adb-bridge | Unicorn ARM64 JNI execution, Flutter AOT introspection, ADB cold trace |
| Browser & Runtime | browser, v8-inspector, mojo-ipc, canvas, webgpu | Camoufox stealth CDP, V8 Turbofan JIT inspection, Mojo IPC sniffing |
| Instrumentation & Memory | memory, process, binary-instrument, native-bridge | AOB scanning, pointer map convergence, Frida dynamic hooks, IDA/Ghidra sync |
| Decompilation & Syntax | wasm, transform, sourcemap, platform | Wasm2C decompilation, AST de-flattening, SourceMap tree recovery |
| Network & Cryptography | network, tls-inspector, proxy, streaming, protocol-analysis, graphql | NSS TLS keylogs, gRPC/Protobuf decoding, MITM proxy, WebSocket monitoring |
| Exploitation & Kernel | syscall-hook, exploit-dev, encoding | Syscall/eBPF monitoring, automated ROP chain discovery, raw Protobuf parsing |
| Orchestration & DevOps | workflow, coordination, extension-registry, tasks, maintenance, instrumentation | Multi-agent coordination board, background tasks, hot-reload plugins |
Read the full tool parameters in the 36 Domains Reference.
.env variables and profile options