Strategy Mosaic and MCP: Governed AI Access to the Semantic Layer

Many organisations now want AI assistants to answer business questions from governed data, not from whatever tables an assistant can reach. Strategy’s response is to expose Strategy Mosaic, its semantic layer, to AI clients through the Model Context Protocol (MCP). Over 2026 this moved from preview to an out-of-the-box capability, and in September 2026 its privileges were split in a way that affects every upgrade.

This article explains what the Mosaic MCP endpoint exposes, how access is controlled, what changed release by release, and the trade-offs for architecture and governance. Product capabilities, editions and entitlements differ by contract, so confirm them for your environment.

What changed in 2026

The Strategy product help page on interacting with agents and Mosaic models using an MCP server gives the release history:

Release	Change
Strategy One (December 2025)	MCP interaction with Strategy agents introduced
Strategy One (February 2026)	Mosaic model access as a preview, enabled in the Collaboration Service configuration
Strategy One (March 2026)	Mosaic access available out of the box
Strategy One (June 2026)	Certification required for Mosaic models; other AI-enabled cubes become accessible
Strategy (July 2026)	Universal agents included in agent listings; dynamic redirect URLs for some clients
Strategy (September 2026)	Privileges split into separate Agent and Mosaic MCP privileges

The naming changed in the same period. The What’s New in Strategy notes for July 2026 state that Strategy One was renamed to Strategy to encompass the BI and Mosaic offerings. In this article, “Strategy Mosaic” means the semantic layer and “Strategy BI” the analytics and reporting product. The October 2026 notes add API-key authentication for agents and MCP clients.

Two MCP endpoints with different jobs

The product help describes two endpoints that share authentication but expose different tools. They can be added to a client as separate connectors.

• Agents MCP, at a path ending /collaboration/mcp/agent, offers list_agents and ask_agent. An AI client delegates the question to a Strategy agent, which answers from its own data, business context and instructions.
• Mosaic MCP, at a path ending /collaboration/mcp/mosaic, offers get_projects, get_models, get_semantics and query. The client reads the semantic model and runs queries against it.

The difference is architectural. With the Agents endpoint, reasoning about the business question stays inside Strategy. With the Mosaic endpoint, the external client reasons and writes the query, and Strategy supplies governed definitions and execution. The help page notes that tool names appear inconsistently in places, so verify them against your release before configuring clients.

Neither choice is universally right. The Agents endpoint suits teams that have already tuned Strategy agents with business context and want a consistent answer wherever the question is asked. The Mosaic endpoint suits technical users and AI-assisted development, where the client needs to see the model’s attributes and metrics and compose its own queries. Many organisations will expose both to different groups.

Supported clients listed on the page include Claude, VS Code Copilot, Amazon Quick Suite, ChatGPT, Copilot Studio and Gemini. Client support depends on each client vendor’s plan and connector features, which change independently of Strategy.

Governed and read-only by design

The Mosaic endpoint is deliberately narrow. According to the product help:

• The query tool runs read-only, Trino-dialect SQL. Write and DDL statements are rejected, so data cannot be changed through MCP.
• Queries use the attributes and metrics defined in each model, so results follow the validated definitions rather than an assistant’s own interpretation of raw tables.
• Only content that has passed Strategy’s certification is exposed. Mosaic models must be published, through certification and viewable by the user. Live and in-memory cubes must also be AI-enabled.
• Access depends on the user’s view permissions and on a separate privilege for each endpoint.

This is the main reason to route AI clients through a semantic layer at all. Metric definitions, joins and security filters are applied once, in one place, for dashboards and assistants alike. The cost is that the semantic model becomes a bottleneck: if a metric is missing or wrong in Mosaic, every assistant inherits the gap.

Certification also becomes an exposure control. The help page advises certifying only the agents you want to expose. In practice that means the certification workflow needs an owner, criteria and a review cadence, because it now decides what AI clients can see.

Privileges, authentication and network prerequisites

The September 2026 privilege split

From the September 2026 release, the product help describes separate privileges:

• Use Agent MCP Server, together with Run AI Bots, for the Agents endpoint.
• Use Mosaic MCP Server for the Mosaic endpoint.

The upgrade behaviour matters. The former Use MCP Server privilege maps to Use Agent MCP Server. Use Mosaic MCP Server is disabled by default and must be granted manually. Teams that used Mosaic through MCP before September 2026 should expect it to stop working for users until the new privilege is granted.

The help page also states that administrators and members of the Platform Support Administrator group cannot use the MCP server. That is a sensible separation of duties, and it means testing must be done with ordinary user accounts that mirror real roles.

Authentication and reachability

MCP access uses OAuth: users sign in to Strategy Library when the client prompts them, and the help page lists the authorisation, token and refresh URLs for manual configuration. From October 2026, the What’s New notes add authentication for agents and MCP clients with API keys.

The prerequisites deserve an early architecture review. The help page states that the environment must be publicly accessible and must have collaboration services enabled. For hosted AI clients, that means the Strategy environment’s endpoint must be reachable from the client vendor’s infrastructure. Organisations with strict network policies, or data-residency requirements, should confirm what is reachable, from where, and through which gateway before a pilot.

API keys suit unattended integrations but are long-lived credentials. Decide who issues them, how they are scoped and rotated, and how their use is attributed in audit records.

Audit, portability and change control

Governance does not end at access. Three capabilities support it, each with limits worth knowing.

• Audit. The Mosaic Sentinel Audit and Compliance module, available since Strategy One (December 2025), keeps an audit trail of Mosaic executions and model changes. It covers semantic clients such as Tableau, Power BI and DBeaver, plus Strategy dashboards. The page does not mention MCP explicitly, and it states that the module does not operate in real time. Confirm that MCP-originated queries appear as you expect before relying on it for compliance.
• Version control. Since March 2026, Mosaic models can be exported and restored as YAML files to support version control. Export needs Read access to all objects in the model and restore needs Control access. Creating new models or modifying objects directly from YAML is not supported.
• Portability. The What’s New notes list import and export of Open Semantic Interchange (OSI) YAML files as a preview in June 2026 and out of the box in July 2026, to use Snowflake OSI data. This can reduce lock-in to a single semantic definition, but test round-trips before you depend on them.

Where this fits in a wider architecture

Mosaic sits between data sources and the tools that consume them. The same governed model can serve Strategy BI dashboards, external BI tools through Strategy’s universal semantic layer connectors, and AI clients through MCP. That is the argument for investing in one semantic model rather than separate definitions per tool.

It also concentrates risk and workload. Query load from AI clients reaches the same model and, through it, the same warehouses. Plan capacity and cost monitoring for that load; the What’s New notes describe Mosaic Sentinel cost intelligence features, which are worth evaluating for this purpose where your edition includes them.

A sensible pilot exposes one well-understood model to a small group, through one client, with audit reviewed weekly. Expand only when answers match the dashboards people already trust.

Decision checklist

• Decide whether AI clients should ask Strategy agents, query Mosaic models directly, or both.
• Confirm the Strategy release in use and the MCP tool names for that release.
• After upgrading to September 2026 or later, grant Use Mosaic MCP Server deliberately, by role.
• Define certification criteria and an owner, since certification now decides AI exposure.
• Review network reachability and data-residency implications of a publicly accessible endpoint.
• Choose OAuth or API keys per use case, with scoping, rotation and attribution rules.
• Verify that Sentinel audit records capture MCP activity as required.
• Keep preview features out of production, as the product help advises.
• Confirm edition, licensing and AI entitlements in your Strategy agreement.

How Bassirah can help

Strategy is a Bassirah strategic technology partner. We help teams design semantic models in Strategy Mosaic, decide how AI clients should reach them, and set up privileges, certification workflows and audit checks before rollout. See our Strategy partner page, and enterprise AI AgentOps for running agents in production. Edition and entitlement details are confirmed in each engagement.