@jshookmcp/jshook

A search-first, profile-aware reverse-engineering workspace for AI agents.

Hook the page, capture the network, deobfuscate the bundle, disassemble the WASM, instrument the process — and let one MCP server keep the whole attack surface in reach without drowning the model in schemas.

English · 中文

What makes jshook different

Most MCP servers for JS analysis expose a handful of hand-rolled tools or wrap a single browser engine. jshook is closer to an operating system for front-end reverse engineering — 34 self-discovered domains, a search-first meta-tool that keeps token cost under control, and runtime recovery that survives broken pages and dropped sessions:

• Search-first, profile-aware. The search profile loads about 3K tokens of tool metadata; the full profile exposes all 723 tools at around 40K tokens. Agents move between them as the task grows — search → workflow → full — instead of drowning in schemas from the first turn.

• Runtime recovery and session isolation. Streamable HTTP sessions restore activated domains, browser attach state, and coverage state after reconnects; per-client browser-side state stays isolated so two agents cannot trample each other's CDP sessions.

• Full-stack browser automation. Chromium and Camoufox via CDP with anti-detection, an explicit-input CAPTCHA solver (no built-in page/feature probing), a self-signed HTTPS interception CA on demand, and HTTP/2 frame building.

• Real reverse engineering, not string searches. WASM disassembly via Binaryen, Frida/Ghidra/IDA bridges, native FFI scanning, hardware breakpoints, PE introspection, GraphQL/Burp Suite proxy bridges, and AST transforms — not a single regex call wrapped as a tool.

• Dynamic extensibility. Hot-reload plugins, declarative workflows, and auto-discovery keep the server growing without a redeploy.

Capability overview

A scan of what's in the box. Each row links to the detailed Capability overview below.

Use cases

Quick start

No global install needed — add to your MCP client config and you're ready.

Claude Desktop / Cursor (claude_desktop_config.json):

{
  "mcpServers": {
    "jshook": {
      "command": "npx",
      "args": ["-y", "@jshookmcp/jshook@latest"],
      "env": {
        "MCP_TOOL_PROFILE": "search",
        "npm_config_omit": "optional"
      }
    }
  }
}

(Windows: use npx.cmd absolute path if npx is not found.)

This lightweight configuration skips optional ONNX, Z3, Binaryen, Camoufox, and Playwright packages. Remove npm_config_omit when those full-profile runtimes are required.

Share one daemon across multiple agents

The default stdio configuration starts one full jshook process per MCP host. To share the embedding model, browser runtime, and caches, start one local Streamable HTTP daemon:

pnpm build
pnpm daemon

Vector search defaults to off for per-client stdio processes and on (lazy-loaded) for the shared HTTP daemon. Set SEARCH_VECTOR_ENABLED=false when lexical search is sufficient.

Then point every MCP client at http://127.0.0.1:3000/mcp using its HTTP/URL server configuration. Each client receives its own MCP session and response route while heavyweight runtime resources remain in one process. Keep the default loopback bind; set MCP_AUTH_TOKEN before exposing the endpoint beyond localhost.

Promote a profile as the task grows

{
  "env": {
    "MCP_TOOL_PROFILE": "search"     // start here, ~3K tokens of metadata
  }
}

Switch MCP_TOOL_PROFILE to workflow once you start chaining composite scripts, or to full when you need every tool. coverage_report shows the active set on demand.

Highlights

• Profile ladder. Start in search (~3K tokens of metadata); promote to workflow when chaining composite scripts; escalate to full only when every tool is actually needed. coverage_report shows what's active on demand.

• Meta tools. describe_tool returns the JSON Schema; call_tool validates arguments before invocation; every tool ships with readOnlyHint / destructiveHint / idempotentHint / openWorldHint.

• Browser automation. Chromium and Camoufox via CDP, attach to existing targets, anti-detection presets, popup/download/permission interceptors, explicit-input CAPTCHA solver, JS/CSS injection at three document phases, persisted coverage across reconnects.

• Network interception. Auto-generated HTTPS interception CA, HTTP/1.1 + HTTP/2 frame building, WebSocket capture, GraphQL helpers, Burp Suite bridge — all on the same MCP tool surface.

• Reverse engineering. Binaryen WASM disassembly, Frida/Ghidra/IDA bridges, hardware breakpoints, native FFI scanning, PE introspection, syscall hooking, AST transforms, source-map reconstruction.

• Session recovery. Streamable HTTP transport restores activated domains, browser attach state, and coverage state after reconnects; browser-side state is isolated per client.

• Plugins and workflows. Drop a directory, get a domain. Write a YAML pipeline, run it as one tool. The registry self-discovers.

Transport and deployment

The server supports two transports out of the box.

Both transports expose the same tool surface. coverage_report shows which domains are activated in each session — long-running sessions restore browser attach state, coverage state, and tool activations across reconnects.

For production deployments see the Security and Production guide.

Recent runtime notes

• HTTP transport now multiplexes independent MCP sessions and restores runtime state after reconnects.

• proxy_start auto-generates a local HTTPS interception CA when needed.

• Browser CAPTCHA solving is now explicit-input driven: pass taskKind, siteKey, imageBase64, callbackName, and responseSelector as needed. Built-in widget/page signature probing is intentionally not used.

Registry snapshot

The built-in surface below is generated from the runtime registry and checked in CI.

• Package version: 0.3.5

• Built-in tools: 723

• Domains: adb-bridge, binary-instrument, boringssl-inspector, browser, canvas, coordination, core, cross-domain, dart-inspector, debugger, encoding, exploit-dev, extension-registry, graphql, instrumentation, maintenance, memory, mojo-ipc, native-bridge, native-emulator, network, platform, process, protocol-analysis, proxy, sourcemap, streaming, syscall-hook, tasks, trace, transform, v8-inspector, wasm, webgpu, workflow

• Note: this snapshot is generated from the runtime registry; do not edit the counts by hand.

View the complete Tool Reference ↗

Architecture

• Runtime registry — domains auto-discovered via manifest.ts; add a domain by creating one file.

• Lazy initialization — handlers instantiated on first call, not at startup.

• BM25 + vector search — search_tools meta-tool with hybrid ranking and adaptive weights.

• MCP ToolAnnotations — every tool carries readOnlyHint / destructiveHint / idempotentHint / openWorldHint.

• Profile ladder — search (~3K tokens) → workflow (composite scripts) → full (all 723 tools).

• Transport symmetry — stdio and Streamable HTTP expose the same surface; sessions are isolated per client.

See the Architecture guide and Configuration reference for the canonical details.

Build from source

Requirements: Node.js 22.12+, pnpm 10.x.

pnpm install
pnpm build
pnpm start           # run the built server from dist/
pnpm dev             # run from source under tsx watch
pnpm check           # metadata check + lint + format check + typecheck + unit tests
pnpm test            # Vitest unit suites
pnpm test:e2e        # end-to-end browser/tooling suites
pnpm daemon          # run the Streamable HTTP daemon after build

Native helpers are bundled via pnpm build; on first run the server may download optional runtimes (ONNX, Z3, Binaryen, Camoufox, Playwright) depending on the profile.

Project stats

License

AGPLv3.