MCP server

The PromptGlass MCP server connects your AI assistant to PromptGlass analytics through the Model Context Protocol. Leaders and analysts can ask questions in plain English, such as “Which teams have the fastest-growing AI spend this quarter?” or “How many Copilot seats have been idle for 60 days?”, and get answers from live data. It is read-only by default and returns pseudonymous, k-anonymous aggregates unless your role allows more.

Install

Claude Code

claude mcp add promptglass -- npx -y @promptglass/mcp

Claude Desktop / Cursor

{
  "mcpServers": {
    "promptglass": {
      "command": "npx",
      "args": ["-y", "@promptglass/mcp"],
      "env": { "PROMPTGLASS_URL": "https://acme.eu.promptglass.ai" }
    }
  }
}

VS Code (Copilot)

{
  "servers": {
    "promptglass": { "type": "stdio", "command": "npx", "args": ["-y", "@promptglass/mcp"] }
  }
}

On first use the server opens a browser for SSO sign-in (OAuth device flow) and uses your PromptGlass role. It never has more access than you do in the console.

Tools

Tool	Returns
get_adoption	Active users, penetration and depth by team/app/period
get_spend	Cost by team, app, model, cost center; trend and forecast
list_idle_seats	Idle seats by tool and team, with reclaimable amount
list_shadow_ai	Unsanctioned apps, user counts, risk scores
get_detections	Detection counts by data class, app and team
get_use_cases	Top clustered use cases and their efficiency signals
explain_policy	What a policy rule does and how often it fired
simulate_policy	Impact of a draft rule on the last N days (no changes applied)

Write tools (apply_policy, sanction_app) are off by default. Turn them on with PROMPTGLASS_MCP_WRITE=true. Even then, each call requires confirmation in the client, and it is recorded in the admin audit log.

Caution

The MCP server never returns prompt content, even to investigators. Use the console for investigations, where access is case-scoped and audited.

Example

You: Which three departments grew AI API spend the most last month, and is it worth it?

Assistant (calls get_spend, get_use_cases): Engineering, Support and Marketing grew the most. In Support, most of the growth is ticket classification on a large model. PromptGlass found no outcome difference with a smaller model, so a routing change would cut that cost. Want me to simulate the routing rule?