CyberCage – AI Security Platform for MCP Servers

TL;DR

TL;DR: CyberCage is an enterprise security platform purpose-built for AI coding assistants and MCP servers — giving teams visibility and control over what data reaches AI tools, which servers are approved, and what’s flowing to LLM providers.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: cybercage.io ← MUST visit and verify
• HN launch thread: news.ycombinator.com/item?id=46235072 ← primary launch
• HN launch thread: news.ycombinator.com/item?id=46717946 ← PII detection launch
• License: Proprietary (enterprise SaaS, no open-source license found on site)

What Is CyberCage?

CyberCage addresses a gap that becomes obvious the moment you ship an AI coding assistant to a team: no built-in visibility into MCP servers, approved tools, or data flow.

MCP (Model Context Protocol) servers let AI tools like Claude Code, Cursor, and GitHub Copilot call external tools at runtime. An MCP server can read files, call APIs, execute code. In enterprise environments, this creates a blind spot — which servers are installed, which employees approved them, what data is being sent where.

The CyberCage founders described the problem directly in their launch post:

“We noticed there wasn’t a good way to manage MCP servers across a team or see what’s actually flowing to LLM providers. Who’s running what? Which tools are approved? What data is going where?”

CyberCage ships three layers of control:

MCP Management

Auto-discovery of MCP servers across the team, with manual or automatic approval workflows. Admins can manage allowed servers org-wide down to the individual tool level. CyberCage also integrates with GitHub’s MCP Catalog to surface a curated, vetted server list.

AI Tool Governance

Controls which AI applications employees can use org-wide. This covers AI IDEs (Claude Code, Cursor, VS Code, Windsurf, Antigravity) and low-code platforms (n8n has a native integration).

Data Leak Prevention

On-device PII detection powered by a local ML model. Scans text and images for:

• SSNs and credit card numbers
• API keys (AWS, GitHub, OpenAI, and custom patterns)
• Phone numbers and other personal data

All detection runs locally — nothing is sent to a remote scanning service. Admins choose the response: log it, block it, or auto-redact before transmission.

Audit and Notifications

Full audit logs with Splunk integration. Notifications via Slack, Teams, Webex, and webhooks. This gives security teams the trail they need for compliance without requiring AI tools to phone home to a third-party scanner.

Setup Workflow

CyberCage is a hosted SaaS platform with an on-device agent component. The exact installation steps require registration, but based on the launch posts and site documentation:

Step 1: Register and Install the Agent

Sign up at cybercage.io and install the CyberCage agent on your workstations. The agent handles local traffic inspection and MCP discovery.

Step 2: Connect Your AI Tools

Authorize CyberCage to monitor configured AI domains. The agent intercepts outbound requests to AI provider endpoints and applies the configured policies.

Step 3: Configure Policies

Set your data policies — which data types to detect, what action to take (log, block, redact). Define approved MCP servers org-wide via the dashboard.

Step 4: Review Audit Logs

Access full audit logs from the CyberCage dashboard. Splunk integration is available for teams with existing SIEM infrastructure.

Deeper Analysis

Why This Matters for AI Tooling

AI coding assistants have access patterns that traditional security tools don’t understand well. A request to an LLM provider might include file contents, API keys pasted into a prompt, or customer data from a database query. Standard DLP tools watch email and web traffic — not AI tool APIs.

CyberCage’s approach of running detection on-device (rather than proxying traffic to a cloud scanner) is architecturally sound for privacy-sensitive environments. The ML model runs locally, and the admin chooses whether to log, block, or redact.

Competitive Landscape

The MCP server ecosystem has grown rapidly, but security tooling has lagged. CyberCage appears to be the first platform purpose-built for MCP governance specifically — most competing products are either general API security tools or generic DLP suites.

The 46717946 launch post focused on PII detection specifically for AI tools — distinguishing between blocking accidental data exposure versus monitoring intentional data sharing.

Limitations

Based on the available information:

• No self-hosted option is mentioned — this is a hosted SaaS product
• BYOLLM (bring your own models for inspection) is listed as “coming soon”
• Browser extensions for AI domain control are also on the roadmap
• The on-device network agent is in private beta

Practical Evaluation Checklist

• Works with Claude Code, Cursor, VS Code, Windsurf, Antigravity
• Native n8n integration for low-code workflows
• On-device PII detection (SSN, credit cards, API keys) — local ML, no cloud scan
• Auto-discovery and approval workflow for MCP servers
• Org-wide MCP server management down to individual tool level
• Audit logs with Splunk integration
• Notifications via Slack, Teams, Webex, webhooks

Security Notes

• PII detection ML model runs entirely on-device
• No data sent to remote services for content scanning
• BYOLLM for inspection is planned but not yet available
• The platform is in active development (multiple HN launches in 2025)

FAQ

Q: Does CyberCage work with self-hosted MCP servers? A: Yes — MCP server discovery covers both hosted and self-hosted installations. The platform discovers servers installed across your team’s workstations.

Q: Is the PII detection model sent to the cloud? A: No. The ML-powered PII detection runs locally on the agent. Nothing is sent off-device for content inspection.

Q: Which AI tools are supported? A: Claude Code, ChatGPT, GitHub Copilot, x.AI, Cursor, VS Code, Windsurf, and Antigravity are explicitly mentioned. n8n has a native integration.

Q: Does this require network proxying? A: The on-device network agent intercepts configured AI domains locally. It does not require routing traffic through a cloud proxy.

Q: Is there an open-source version? A: The founders mentioned plans to open-source CyberSmol v1.0 — a small model fine-tuned for AI threat detection. As of the latest launch post, this has not shipped yet.

Conclusion

CyberCage fills a real gap in the AI tooling stack — teams adopting MCP-based AI coding assistants have had no standard way to govern what those servers can access or what data leaves the network. CyberCage brings enterprise security primitives (discovery, approval, DLP, audit) into the MCP era.

The on-device PII detection is the most practical feature for day-to-day use — it catches accidental API key or SSN exposure before it reaches an AI provider, without sending content to a third-party scanner. Watch for the BYOLLM and browser extension releases to expand coverage.

Source last checked: 2026-06-25 (HN launches 46235072, 46717946, 46703915)