Verification Tools as an MCP Server

The calculator implements exactly the models published in Measuring Verification Debt and What Verification Debt Costs - same formulas, same thresholds, same sources. Every answer ends with its calculation path, because that is the point: claims should carry their evidence.

This page is human documentation. The protocol endpoint is /api/mcp. Machine-readable discovery is available at /.well-known/mcp, /.well-known/mcp.json and /.well-known/mcp/server-card.json. These JSON files describe only the public read-only surface: documentation, legal/privacy links, the stateless MCP endpoint, and the tool names. They do not expose admin, lead, customer, private briefing, analytics, GEO admin, repository-private or roadmap-private data.

curl -sS https://realitygraph.dev/api/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  --data '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "initialize",
    "params": {
      "protocolVersion": "2025-11-25",
      "capabilities": {},
      "clientInfo": { "name": "smoke-test", "version": "0.0.0" }
    }
  }'

A browser GET to /api/mcp returns 405 by design; this stateless MCP endpoint accepts POSTed JSON-RPC requests.

Custom connectors are available from the Free plan (one custom connector; unlimited on Pro and above). In Claude, open Settings → Connectors → Add custom connector, paste the endpoint URL, and add it - no OAuth fields needed. Connectors added on the web also sync to the mobile apps.

Settings → Connectors → Add custom connector

Name:  Reality Graph Verification Tools
URL:   https://realitygraph.dev/api/mcp

For Claude Code, one command is enough:

claude mcp add --transport http reality-graph https://realitygraph.dev/api/mcp

Custom MCP connectors require developer mode (Plus, Pro, Business, Enterprise or Edu - on the web). Enable it under Settings → Apps & Connectors → Advanced settings → Developer mode, then create the connector:

Settings → Connectors → Create

Name:            Reality Graph Verification Tools
MCP server URL:  https://realitygraph.dev/api/mcp
Authentication:  No authentication

Because the server exposes tools named search and fetch, it also works as a source in ChatGPT’s research-style connector mode, not just in developer mode.

Perplexity (Pro, Max, Enterprise): Settings → Connectors → Add custom connector → select Remote, paste the URL, set Authentication to None and Transport to Streamable HTTP.

Cursor - add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "reality-graph": {
      "url": "https://realitygraph.dev/api/mcp"
    }
  }
}

VS Code (GitHub Copilot) - run MCP: Add Server from the command palette, or add to mcp.json:

{
  "servers": {
    "reality-graph": {
      "type": "http",
      "url": "https://realitygraph.dev/api/mcp"
    }
  }
}

The endpoint is operated from Germany under GDPR by the operator named in the legal notice; details are in the privacy policy. The server follows the MCP specification’s stateless Streamable HTTP transport; tool descriptions are versioned in the open and never change at runtime. If you find a security issue, contact us via the contact page.

The tools are designed as one loop around an AI coding run - the same loop the method articles describe, executable by hand and for free:

1. Before: get_task_contract_template → fill it in → lint_task_spec / validate_task_contract until it passes.
2. Run: your AI coding tool of choice does the work - the contract travels with the task.
3. After: get_verification_report_template → record what was checked, what was skipped, and who decided.

Everything here is the manual edition of the loop Reality Graph runs locally as a product (private beta): evidence reports per run, with the same discipline and none of the typing. If the manual loop earns its keep in your team, that is the signal - early access is open.

Beside the remote server, this site registers its tools in the browser via the emerging WebMCP standard (Chrome 149+ origin trial, document.modelContext): an in-browser AI agent visiting realitygraph.dev gets the same 10 callable, read-only tools instead of guessing at the UI, including lint_task_spec, check_verification_debt and search + fetch for article discovery and reading. Same implementation, same privacy posture - the page tools call the same stateless endpoint.

Reality Graph’s knowledge base - verification debt, the verification gap statistics, the glossary - is written to be quoted, with self-contained definitions and named sources. An MCP server is the shortest path from that knowledge to the place developers actually ask questions: their AI assistant. The product itself, a local-first verification layer for AI coding, stays local by design - this server only serves what is already public. A coding agent working inside a repository talks to a different one: a local, read-only server that puts its refusals in writing. Same protocol, different machine, different data.