Every number here comes from production: each model call, search and price lookup is logged with how long it took and whether it worked. Nothing about a patient or a bill is shown.

 

Loading live numbers…

Checks that keep the agents honest

Fixed tests. Each one can be rerun from the repo.

  • 24/24planted billing errors caught, no false alarms
  • 80/81Apify prices matched hand-checked prices
  • 340/470California hospitals with line-by-line price checks
  • 470hospital policies read in the statewide audit

Check it yourself over MCP

Our MCP server gives these same numbers to your own AI client, read only. Add it to VS Code (GitHub Copilot), Claude or any MCP client, then ask “How reliable are Recount's agents this week?”

Server addresshttps://recount-nu.vercel.app/api/mcp
VS Code: .vscode/mcp.json
{
  "servers": {
    "recount": {
      "type": "http",
      "url": "https://recount-nu.vercel.app/api/mcp"
    }
  }
}
Claude Code
claude mcp add --transport http recount https://recount-nu.vercel.app/api/mcp
Tools
  • get_agent_metrics Runs, success rate, timing and quality per agent
  • get_agent_timeseries The same, in 15-minute steps
  • get_recent_agent_runs The 15 latest runs
  • get_validation_results The fixed test results above, with sources