USE CASES / CONTEXT FOR AI (MCP)
Context for AI: give every agent the metrics your business already trusts.
The AtScale MCP server gives AI assistants and agents certified metrics by name, so they don’t have to guess from raw tables.
An LLM with access to your warehouse still has to work out which tables, columns and calculations a question needs. AtScale gives it the context it’s missing: certified metrics with plain-language descriptions, the rules for using them, and an engine that computes the answer. Every AI client and agent gets it through one MCP server.
The challenge
Agents writing their own SQL get some numbers wrong, and they don’t wait for anyone to check before they act.
Context in prompts and markdown files goes stale, and it doesn’t carry over to other models.
You’d need a separate integration for each AI tool.
What you get
Agents ask for a certified metric by name and get a computed answer.
One semantic model serves every AI client, with each agent signed in as itself.
THE CHALLENGE
More context helps, but the AI still does the math.
Teams try to give AI context by pasting definitions into prompts, writing skills, or pointing it at a catalog. That helps the model find the right table, but it’s still writing the SQL and doing the calculation, and it may do both differently on the next run. Anthropic’s data team saw accuracy drift from about 95% to about 65% within a month when their hand-written context fell behind (Why skills alone aren’t enough).
HOW IT WORKS
Four calls from question to record.
- Find the metric. The agent searches certified metrics and dimensions by their business descriptions with
list_models,explore_columnsandfocus_columns. - Learn the rules.
get_sml_skillstells it how to query the model, including terms with more than one meaning. - Get the answer.
run_queryasks for the metric, and ACE computes it in your data platform. - Keep the record.
get_outbound_queriesreturns the exact SQL behind any answer.
People asking in a chat go through AtScale Ava, which checks what they’re asking for before anything runs.
WHERE IT WORKS
Any AI client or agent that speaks MCP.
Claude, ChatGPT, Microsoft Copilot, Grok and Codex can all query a published model, and so can agents built with Claude’s and OpenAI’s agent SDKs, LangGraph or Google’s Agent Development Kit. Each agent signs in through OAuth with its own identity, and row- and column-level security applies to it the same way it applies to a person. Every call is logged.
THE PROOF
Twice as accurate with the right context.
On BIRD-Interact, models that reached AtScale over MCP were twice as accurate as the same models writing SQL on their own. AtScale’s results are preliminary.
On BIRD-Interact, models that reached AtScale over MCP were twice as accurate as the same models writing SQL on their own. AtScale’s results are preliminary. Blue Yonder connects 600 semantic objects to its internal AI agents through the AtScale MCP server, so its agents and dashboards use the same metrics.
FAQ
Questions about context for AI.
Does this replace our data catalog?
No. A catalog documents where data lives, and ACE, AtScale’s AI Computation Engine, computes the number. CoModel, coming this fall, will read your catalog when it drafts the model. Why a catalog isn’t enough.
Which AI tools can connect?
Claude, ChatGPT and Google Gemini, plus any client or framework that supports remote MCP servers. Databricks agents can connect through AtScale’s Databricks Marketplace listing.
Does it work with agents we build ourselves?
Yes. If it can call a remote MCP server, it can use AtScale. That includes agents built with Claude’s and OpenAI’s agent SDKs, LangGraph or Google’s Agent Development Kit.
How do agents sign in?
Through OAuth, each with its own identity. Row- and column-level security applies to an agent the same way it applies to a person, and every call is logged.
Can an agent run any SQL it likes?
No. The MCP tools are read-only, AtScale checks every request against the model, and the agent only gets metrics it’s allowed to see.
What about agents on other platforms?
AtScale plans to add the Agent2Agent (A2A) protocol for them. Until then, any agent that supports remote MCP can connect.
Book a workshop.
The workshop takes thirty days from kickoff to results.
- Bring us your most controversial metric and the business questions your team and your agents ask about it every week.
- We benchmark your AI on them, with AtScale and without it.
- You keep the semantic model we built to run the test.