NetSpectre WebMCP Challenge Edition
by DoobyDev
README.md
# NetSpectre WebMCP Challenge Edition
NetSpectre Challenge Edition is a standalone, browser-based demonstration of
human-agent collaboration in a defensive cyber investigation. It was created
for the 2026 WebMCP Challenge and uses synthetic data exclusively.
This repository is **not** the NetSpectre commercial desktop product. It does
not contain NetSpectre's private investigation engine, production data,
customer information, credentials, or proprietary detection logic.
**Live demonstration:** https://netspectre-webmcp.stuhall44.chatgpt.site
## What the demonstration shows
The included fictional case concerns a supplier-impersonation and payment-
diversion attempt. A person and their agent share the same visible case state:
1. The agent obtains a concise case briefing.
2. It reviews preserved evidence metadata without opening or executing files.
3. It explains prioritised signals and their evidence links.
4. It prepares a finding as a draft.
5. The investigator explicitly reviews and approves the finding.
6. The agent generates a summary that reflects the recorded human decision.
## WebMCP tools
| Tool | Behaviour |
| --- | --- |
| `get_case_briefing` | Reads the active case, risk, confidence, and state. |
| `list_case_evidence` | Reads safe metadata for synthetic evidence. |
| `list_prioritised_signals` | Reads explainable signals above an optional confidence threshold. |
| `draft_investigation_finding` | Creates a visible draft that still requires human approval. |
| `generate_incident_summary` | Reads a summary of the current visible review state. |
The app feature-detects `document.modelContext.registerTool`, so its normal
interface continues to work in browsers without WebMCP support.
## Safety properties
- Synthetic evidence only.
- No live scanning, network requests, shell execution, or arbitrary file access.
- Evidence metadata is read-only and no suspicious material is executed.
- Agent-created findings remain drafts until a person approves them visibly.
- Tool inputs use narrow JSON schemas.
- Tool activity is reflected in the visible activity record.
- Outputs support rather than replace professional human judgement.
## Run locally
Use Node.js 22.13 or newer.
```sh
npm install
npm run dev
```
Open the printed local URL. For WebMCP discovery, use ChatGPT's built-in
browser or a supported Chrome build with WebMCP testing enabled.
## Build
```sh
npm run build
```
## Challenge work statement
This standalone WebMCP demonstration was created during the WebMCP Challenge
submission period. NetSpectre existed previously as a private Windows-first
desktop product; judges should evaluate the work in this repository as the new
challenge extension and demonstration.
## Licence boundary
The MIT licence in this repository applies only to the files contained in this
challenge-edition repository. It does not grant rights to the separate
NetSpectre commercial desktop application, its private source, data, brands,
or other proprietary materials.
This server cannot be deployed
Maintenance
ActivityMaintained
ResponsivenessNo issues