Bloomberg Enterprise MCP checks entitlements on the server

Bloomberg launched Bloomberg Enterprise MCP on September 29. Client AI agents can now find and pull licensed Bloomberg data across more than 100 million securities and 50,000-plus fields through one Model Context Protocol endpoint. The notable part is not the data. It is where the rules live: Bloomberg checks each firm's entitlements on its own server before it returns a single row. That is the right design for any business that exposes data to agents, including yours.

What actually happened

Per Bloomberg's announcement, Enterprise MCP is a hosted access layer for Data License Plus (DL+) clients. It pairs the data with machine-readable metadata and semantic context, plus workflow "Skills," so an agent can discover the right dataset in natural language before it asks for it.

Coverage includes pricing, reference data, company fundamentals, economics and alternative data. Bloomberg also ships task-level tools: outlier detection, price deviation monitoring, sanctions screening, point-in-time universe retrieval and corporate-action adjustments.

The controls are specific. Bloomberg validates entitlements before it returns data, resolves requests against an operational datastore, and caps how many securities and fields each tool call can request. Clients bring their own agents; Bloomberg hosts the server. Real-time data and a Terminal version are on the roadmap.

Tony McManus, Bloomberg's global head of enterprise data, said the bottleneck for financial firms has moved from model capability to data readiness.

Why the Bloomberg Enterprise MCP design matters for your business

You will not buy a DL+ license. But you will soon expose your own data — inventory, order status, pricing, a customer's account — to somebody's agent. Bloomberg's pattern is the one to copy:

Enforce permissions on the server. Never trust the agent to request only what it may see. Check the caller's scope on every call.

Cap every call. Per-call limits on rows and fields stop a runaway loop from dumping your catalog.

Ship the metadata. Field descriptions and units are what let an agent pick the right tool. Without them, it guesses.

Build task tools, not raw tables. A "check order status" tool is safer and more useful than open read access to the orders table.

MCP is now how a decades-old data company sells to AI. The same interface is available to a 10-person shop.

Sources: Bloomberg via PR Newswire.