The Jefferies affiliated analytics shop has wrapped its 1,440 KPI institutional data estate in an MCP server, letting AI copilots and agents query its star schema data layer directly — a structural bet that the next workflow bottleneck is not the model, but the data plumbing…
Wall Street's trading floors have spent two years asking whether AI can replace analysts. The harder question, it turns out, is whether the data underneath those analysts can keep up with the agents.
On June 2, 2026, M Science — the Jefferies-affiliated analytics and data firm — launched a Unified Data Model and an MCP (Model Context Protocol) server that exposes its institutional data estate to OpenAI's ChatGPT, Anthropic's Claude, and clients' own internal AI copilots. The move is plumbing-level rather than splashy, but it points at where the next workflow bottleneck is showing up: not in the model, but in the data layer that feeds it.
M Science's announcement is a product, not a vision. The Unified Data Model organizes more than 1,440 KPIs across 1,400+ companies into a star-schema warehouse, with change-data-capture (CDC) pipelines that keep point-in-time and back-test workflows accurate. The MCP server sits on top, translating the data layer into a protocol that AI agents and copilots can query directly.
In practical terms, an analyst running an internal Claude-based research assistant can now ask the model to pull M Science's curated KPIs — alt-data signals, fundamentals, supply-chain indicators — without an engineer building a custom integration. The same protocol extension is what powers "Maddie," M Science's in-house AI copilot, and is now being handed to customers. CEO Michael Marrale framed the launch as "the infrastructure clients need to operationalize [data] at scale."
The architecture matters more than the marketing. A star schema with CDC means the data has structure agents can reason over, plus a time-aware refresh path so a back-test doesn't accidentally use future revisions. Most institutional datasets are still flat files, REST endpoints, or Excel drops — formats that work for humans with a Bloomberg Terminal, and that break the moment you try to hand them to a model.
MCP is no longer a research curiosity. At the MCP Dev Summit North America 2026 in New York, the Agentic AI Foundation (AAIF), housed under the Linux Foundation, reported that MCP sees roughly 110 million SDK downloads per month. OpenAI's agent SDK and LangChain now pull MCP in as a default dependency. Behind corporate firewalls, the same recap notes, MCP is being wired to Salesforce, Jira, Snowflake, internal wikis, and HR systems — a sign that the protocol is hardening into the kind of cross-vendor substrate that HTTP became for the web.
AAIF, which adopted MCP, Goose, and AGENTS.md as flagship projects and rolled out a three-stage project lifecycle, said it counts 170 member organizations less than four months after standing up — a faster ramp than CNCF managed at the same stage. The Linux Foundation has called MCP "the linux of agents." MCP co-creator David Soria Parra of Anthropic summarized the shift at the summit: the bottleneck is moving from the model to the data layer underneath it.
That is the shift M Science is positioning for.
A Bloomberg Terminal — which NeuGroup pegged at $31,980 per year for a single seat in 2025, with multi-Terminal seats running $28,320 and a 6.5% increase tied to weighted global inflation — is, among other things, a beautifully packaged routing system for institutional data. (The 2026 list price has not been confirmed as of this writing.) M Science's MCP server does not replace that. What it does is let an analyst ask an AI agent for a KPI directly, then chain that query into a model that drafts a memo, runs a scenario, or populates a dashboard. The terminal still has the chat, the news, the IB messaging. The data layer it aggregates is now being unbundled, one MCP server at a time.
That is the constructive read. The skeptical read is just as real: this is a vendor press release. M Science did not disclose adoption metrics, named customers, or third-party validation. The same NeuGroup post that documented Bloomberg's pricing also noted that at least one member has moved some employees off Bloomberg to a cheaper alternative, while another said they are "evaluating other options, but it's likely we remain with Bloomberg for the foreseeable future." MCP is real, but it is also early. The summit that crowned it drew 1,200 attendees — a doubling, but still a room, not a market.
Three things are worth watching as the institutional data layer goes agent-native.
First, whether the other major vendors — S&P Global, FactSet, LSEG/Refinitiv, MSCI, and Bloomberg itself — ship or announce their own MCP-compatible servers. The story gets stronger if it is industry-wide plumbing, and weaker if it is one Jefferies-affiliated shop with a marketing lead.
Second, whether buy-side firms actually let agents query vendor data directly, or wrap the same queries in governance layers that look a lot like the old vendor relationships. MCP is a protocol, not a procurement decision.
Third, what an analyst can do tomorrow that they could not do last quarter. The honest answer today is: pull a curated KPI into a Claude or ChatGPT session, with CDC-clean point-in-time data, and chain it into a model-driven workflow. That is real. It is also incremental. The bigger claim — that the keyboard-and-terminal interface gives way to a chat-with-your-data interface — is a structural bet, not a delivered outcome.
M Science has staked a position on which side of that bet comes first. The data plumbing, the company is saying, is the new model.