S&P Global Energy opens commodity data to agents on preview MCP servers

S&P Global Energy has opened its commodity data estate to outside AI agents by exposing each curated Databricks Genie Agent as a managed MCP server, according to a case study Databricks published on 25 September. The servers carrying that traffic are still marked Public Preview in Databricks’ own documentation, and the write-up reports no measured result beyond a vice president saying work that once took a full development cycle “now takes days”.

The mechanism is curation rather than code, which is the point of it. Subject matter experts pick the tables for a business domain, then create one Genie Agent per dataset group instead of one per commodity. Liquefied natural gas alone gets seven, covering assets and contracts, cargo, tenders, outages, supply and demand, netbacks, and prices. Non-Databricks tables arrive through Lakehouse Federation connectors, which the post says needed no new ETL job.

Each of those agents then becomes an MCP endpoint automatically, at the form https://<workspace-hostname>/api/2.0/mcp/genie/{genie_space_id}. The post describes a two tool surface per agent: genie_query_space submits a natural language question, and genie_poll_response retrieves the answer with its generated SQL once the warehouse finishes. That ask then poll pattern runs asynchronously, which suits agents that fan a question out and wait.

Genie Agents let our domain experts productize their knowledge of the data directly. What used to take a full development cycle now takes days, and every answer stays inside our governance boundary.

Priyanka John, Vice President, S&P Global Energy, via Databricks

Real questions cross those boundaries, so the team put a FastMCP proxy on top. One composite endpoint per commodity mounts that commodity’s group level Genie servers behind name-spaced tools, so a client configures one server rather than a dozen. The post’s example of a cross-group question is how recent outages at Sabine Pass affected cargo premiums into Asia. The agent’s model picks which group Genie to route to, or fans out and synthesises.

The catch is that what the case study leaves out sits in the documentation it points at. The managed MCP page, last updated on 21 September, carries a Public Preview banner, says on-behalf-of user authentication needs the genie OAuth scope, and states that Genie Agents bill on serverless SQL compute pricing. That last line matters because every customer question is a warehouse query, and the post’s own architecture invites external customers to connect their own clients.

Claim in the case study	What the documentation adds
Nothing to deploy, nothing to host	Managed MCP servers are in Public Preview
Authentication is handled by the platform	On-behalf-of user auth needs the genie OAuth scope
No figures given for cost	Genie Agents use serverless SQL compute pricing
No figures given for latency	Listing tools runs 300 to 400ms per HTTP mounted server, against 1 to 2ms for a local tool

The latency figure comes from FastMCP’s composition page, which warns that a parent server’s list_tools() is affected by the performance of every server mounted under it. It suggests caching or limiting mounting depth when low latency is critical. S&P Global Energy’s own listed lesson is to measure trust, not just latency, and the team tracked how often its experts agreed with Genie’s generated SQL. Neither number is published.

One smaller mismatch is worth flagging for anyone copying the code. The published snippet uses FastMCP.as_proxy() and a prefix argument to mount(), while the current docs document create_proxy() and a namespace argument. The post does label the snippet illustrative and tells you to adapt it to your FastMCP version. FastMCP is moving fast enough to justify that: the PyPI package index records 4.0.10 as the current version, uploaded on 25 September, the same day the case study went up.

The wider bet is that MCP, an open protocol for connecting models to external data and tools, is stable enough to be a customer-facing contract. That is a bigger commitment than wiring up an internal assistant, because it moves an enterprise AI deployment from a pilot into a customer-facing product. Databricks now also recommends starting with its workspace-wide Genie One MCP server for analytics, which is not the per-agent server this architecture composes.

What to watch is whether the preview label comes off before customers depend on it, and whether anyone publishes accuracy numbers from the Genie Agent Benchmarks the post describes. Those benchmarks score agents against verified answers and can be rerun after a change, which is the honest way to hold a text to SQL product steady. Until then, this reads as a well-documented pattern rather than a measured outcome. Even so, the governance answer is the clearest part of it: Unity Catalog permissions decide what each question can reach, which is the same problem Ethyca built a whole product around.