接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令? 兩種安裝方式的差異
來自發布者
Interactive CVE intelligence map for desktop: force-directed Vendor↔CVE graph plus a local AI agent (MCP + OpenRouter) that drafts mitigation plans and keeps versioned, resumable chat reports. Built on Tauri 2, Rust, Python, Vanilla JS.
👁️ A desktop-native CVE intelligence map — interactive force-directed graph of vendor↔CVE relationships wired to a local AI agent that drafts mitigation plans, maintains a versioned chat history, and resumes conversations across sessions. Actively exploited CVEs are flagged straight from the CISA KEV catalog.
Built on Tauri 2 (Rust shell), Python (async AI + MCP client), and Vanilla JS with vis-network. Local-first, no telemetry, no cloud lock-in.
git clone https://github.com/sleepti3ht/Panopticon.git
cd Panopticon
# Python backend
cd panopticon-python
python -m venv venv
venv\Scripts\activate # Windows
source venv/bin/activate # macOS / Linux
pip install -r requirements.txt
cp .env.example .env # add your OPENROUTER_API_KEY
# Frontend + Tauri
cd ..
npm install
npm run tauri dev
First launch auto-creates
panopticon.db(SQLite) with the schema.
cd panopticon-python
python seed_mock.py # 50 synthetic CVEs — instant, no API keys
python ingestor.py 4000 90 # real NVD feed: last 90 days, up to 4000 CVEs with CPE
The ingestor uses a rolling window counted from request time and auto-chunks requests to respect the NVD 120-day-per-query limit. With a free NVD_API_KEY in .env it runs ~10× faster (50 vs 5 requests per 30s).
.env)🕸 Force-directed threat graph
vis-network with forceAtlas2Based solver, auto-stabilization and CPU-off after 1500 iterationsnetwork.setOptions() — no re-render🔥 CISA KEV awareness
ACTIVELY EXPLOITED row with patch due date in the CVE details panel🤖 Local AI mitigation agent
AVAILABLE_MODELS (tested: nemotron-3.5-lightning, gemma-4-31b-it, qwen-3.8 family)💬 Versioned chat with memory
2/3)isProcessing guard + finally-unblock)📚 Reports — resumable conversations
save_report.py purge cleans legacy rows with invalid JSON payloads⚙️ Desktop shell
localStorage for API keys and export pathVulnerability triage today is a browser tab soup: NVD, vendor advisories, CISA KEV, ChatGPT, internal runbooks — all in separate windows. The analyst context-switches constantly and the mental model of "this CVE, this vendor, these affected products, these mitigations" never lives in one place.
Panopticon collapses that into a single native surface:
┌─────────────────────────────────────────────────────────────────┐
│ Vanilla JS (Vite + vis-network) │
│ │
│ invoke("chat_with_agent", { cveId, messages, model }) │
└────────────────────┬────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ Tauri 2 · Rust command handlers │
│ │
│ async fn chat_with_agent(cve_id, messages, model) │
└────────────────────┬────────────────────────────────────────────┘
│ spawn python.exe · history via stdin
▼
┌─────────────────────────────────────────────────────────────────┐
│ Python asyncio layer │
│ │
│ ai_agent.py ──► MCP client ──► mcp_server.py (CVE context) │
│ │ │ │
│ │ ├─► SQLite (local CVE DB) │
│ │ └─► CISA KEV catalog │
│ │ + NVD live fallback │
│ └──► OpenRouter API + SQLite (chat_reports) │
└─────────────────────────────────────────────────────────────────┘
The Rust layer is deliberately thin: it owns the window, routes IPC, and spawns Python processes. All heavy logic (LLM orchestration, graph building, search, persistence) lives in Python, which is easier to iterate on for ML-adjacent work.
Panopticon/ # repo root = desktop project
├── src/ # Vanilla JS + CSS
│ ├── main.js # UI, graph, chat, panels, reports
│ └── styles.css # Dark/light theme (purple accents)
├── src-tauri/ # Rust shell + IPC commands
│ ├── src/lib.rs # Thin handlers: routing + process spawn
│ └── tauri.conf.json # Window defaults, permissions
── panopticon-python/ # Backend
│ ├── ai_agent.py # LLM orchestration + MCP client
│ ├── mcp_server.py # MCP tools: CVE context, KEV, CWE stats
│ ├── graph_builder.py # Vendor↔CVE graph export + KEV flags
│ ├── global_search.py # Cross-table search
│ ├── get_vendors.py # Vendor list with CVE counts
│ ├── save_report.py # chat_reports CRUD + migrations
│ ├── ingestor.py # NVD → SQLite rolling-window pipeline
│ ├── db.py # Schema + atomic inserts
│ ├── config.py # Env-driven constants
│ ├── utils.py # Secret masking + degenerate-repetition helpers
│ ├── seed_mock.py # 50 synthetic CVEs for demos
│ └── panopticon.db # (auto-created, gitignored)
├── docs/
│ └── img/ # README + article screenshots
├── index.html
├── vite.config.js
└── package.json
json_repair integration for malformed LLM outputslocalStorage API keys).panopticon export/import for team handoffsRead the full architectural breakdown on dev.to: Building Panopticon: A Local-First CVE Intelligence Map with Tauri 2 + Rust + Python
Covers sampling penalties for LLM repetition collapse, stdin transport bypassing Windows CLI limits, CISA KEV projection bugs, and why vanilla JS won over React for this tool.
Built something with Panopticon's architecture? Open a PR adding it to this list — one line with a link and a short description.
MIT
Interactive CVE intelligence map for desktop: force-directed Vendor↔CVE graph plus a local AI agent (MCP + OpenRouter) that drafts mitigation plans and keeps versioned, resumable chat reports. Built o
總覽
Interactive CVE intelligence map for desktop: force-directed Vendor↔CVE graph plus a local AI agent (MCP + OpenRouter) that drafts mitigation plans and keeps versioned, resumable chat reports. Built on Tauri 2, Rust, Python, Vanilla JS.
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式,而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Panopticon 透過 stdio 在本機執行。
上次檢查時間:2026-10-28。 Interactive CVE intelligence map for desktop: force-directed Vendor↔CVE graph plus a local AI agent (MCP + OpenRouter) that drafts mitigation plans and keeps versioned, resumable chat reports. Built on Tauri 2, Rust, Python, Vanilla JS.