Vr Org MCP Server

vr-org-mcp

Read-only Model Context Protocol server for VR.org, a dedicated VR / AR / XR news publication and aggregator.

It gives any MCP-compatible agent (Claude Desktop, Claude Code, Cursor, Continue, and others) one-call access to live VR, AR, and XR news, VR.org's original editorial (including full article text), the VR/AR/XR events calendar, curated headset deals, buyer-guide answers, and top-game and top-app lists.

Eleven tools, five resources, and three prompts. Zero keys. Zero writes. Zero payments.

Install

Run it directly with npx (no global install needed):

npx vr-org-mcp

Claude Desktop

One-click: download vr-org-mcp-0.3.2.mcpb and double-click it to install into Claude Desktop. Or configure manually:

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "vr-org": {
      "command": "npx",
      "args": ["-y", "vr-org-mcp"]
    }
  }
}

Claude Code

claude mcp add vr-org -- npx -y vr-org-mcp

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "vr-org": {
      "command": "npx",
      "args": ["-y", "vr-org-mcp"]
    }
  }
}

Hosted remote endpoint (no install)

VR.org also runs the same tools as a remote server over MCP's streamable-HTTP transport, so web clients (ChatGPT connectors, Claude.ai connectors) can use it with no local install:

https://vr.org/mcp

Human setup walkthrough for every client: vr.org/connect.

Tools

Resources

Browsable MCP resources an app can attach as context:

Prompts

How it works

Every tool composes VR.org's public JSON API (https://vr.org/api/*) into a single agent-friendly response. The server is a thin proxy: it holds no secrets, writes nothing, and cannot move money.

Threat model

VR.org's editorial is controlled, but the live feed also carries third-party RSS headlines. To keep a malicious or compromised upstream headline from manipulating the calling model, every tool output is:

1. Sanitized. Control characters and zero-width / direction-override characters are stripped from every string.
2. Capped. Serialized responses are limited to 50 KB so a large payload cannot flood the agent's context window.

Inputs are validated before any outbound request and every free-text parameter is length-capped at the schema layer. Errors are returned as structured, non-echoing objects rather than raw stack traces; on the rare path where an error string reaches the caller it runs through the same output scrub as a tool result, with any credential-shaped substring redacted and the text capped, and is flagged with isError.

Configuration

The API base URL is fixed to https://vr.org in source and cannot be overridden, so the server can only ever talk to VR.org.

Development

npm install
npm run dev        # run from source over stdio
npm run build      # compile to dist/
npm test           # run the offline test suite
npm run typecheck  # type-check without emitting

Changelog

0.4.0

Content provenance. Feed results now carry a provenance field on every item, either vr_org_editorial (written and edited by VR.org) or third_party_feed (a headline or snippet relayed verbatim from an outside publisher). Any response containing relayed text also carries a content_notice saying to treat that text as data rather than instructions, and the vrorg://news/latest resource carries the same notice in prose. This is additive metadata, so existing fields and existing consumers are unaffected.

The point is to let a calling client tell which text VR.org actually stands behind. It follows the August 2026 GhostSplice research on splitting instructions across MCP channels, whose core mitigation is that clients should treat server output as data. This server is not a vector for that attack (it never requests sampling, its tool descriptions are static literals, and every value that can flow back in as a tool argument is validated against a strict allowlist), but it does relay text it did not write, and now it says so.

Also bumps @modelcontextprotocol/sdk to 1.30.0 and pins four transitive dependencies of the SDK's HTTP transport that carried advisories. This server is stdio-only and never loads that transport, so the pins are hygiene rather than an exposure fix. npm audit --omit=dev reports zero vulnerabilities.

0.3.2

Error-path hardening: tool errors now pass through the same output scrub as tool results, with secrets redacted and error text capped. All free-text parameters gain schema-level max lengths. No behavior change for valid inputs.

License

MIT. A VR.org project.