Preset MCP: 25 Tools, and the One Nobody Else Ships
Every BI vendor has an AI story now. I've read a lot of them, and they mostly rhyme: ask a question in plain English, get a chart back. Fine. Useful, even. But it's the easy half of the problem, and after a while the demos all start to blur together.
The question I'd actually want answered as a buyer is less flattering: once your AI has read my data and drawn me a chart, what else can it do? Can it pick up the next part of the job, or does the work evaporate the second I close the tab?
That's the real line, and it's a bigger one than it sounds. An AI that can only read your data is a very fancy search box. An AI that can write back to it is doing the work alongside you, and leaving something behind when it's done.
MCP, quickly
Model Context Protocol is an open standard that lets an AI client — Claude, ChatGPT, Cursor, whatever you've got — talk to outside tools in a structured, secure way. It's a universal adapter between the model and the systems it needs to touch.
For BI, that means an agent can do more than turn your question into SQL. It can discover which datasets exist, read their schema, query them through governed metrics, and (this is the part I care about) create new data assets that outlive the conversation. We wrote about the promise of MCP-powered data workflows a while back. This post is about what changes once the tools can write, not just read.
We shipped Preset's MCP server in April with 20 tools. In May we took it to 25, across five domains: system health, discovery, read operations, chart and dashboard management, and dataset/SQL operations.
It's those last two, the dataset and SQL operations, that I want to talk about. They're the ones nobody else ships.
Most BI MCP servers are read-only. Ours isn't.
I went looking at every major BI vendor's MCP implementation I could find: Tableau (16 tools, GA), Power BI (20+ tool categories, Public Preview), Looker (managed, Preview), ThoughtSpot (GA with Spotter 3), Sigma (GA), Metabase (GA), Hex (Beta), Omni, and Lightdash. I clicked through all of them. And with very few exceptions, they come down to the same thing: a handful of read-only tools and a press release. Find a dashboard. Run a report. Summarize a chart.
Preset's server does all of that too. It also does the thing none of the others do: it creates.
create_virtual_dataset
This one lets an AI agent save a SQL query as a reusable virtual dataset in Preset. Not a throwaway result set. A governed, named dataset that other people, other dashboards, and future AI sessions can build on.
Most AI analytics today is disposable. You ask, you get an answer, and then the work is gone. The next person who needs the same number starts over from a blank prompt. That has always struck me as a waste.
create_virtual_dataset is the fix. An analyst working in Claude or Cursor can chase a question, refine the SQL, and then save the result as a first-class dataset in Preset, with column descriptions, metrics, and governance intact. From that moment it's available to every other user and every future AI session. The conversation ended; the asset didn't.
No other BI vendor's MCP server does this today. Tableau's reads datasources and workbooks. ThoughtSpot's spins up analysis sessions for conversational reasoning. Looker's queries governed models. All useful. All read-only. Preset is the only one where the AI actively adds to your data layer.
query_dataset
This one lets an agent query datasets through saved metrics, calculated columns, and dimensions instead of raw SQL. It works through the semantic layer, respecting the definitions and business logic your team already argued about and agreed on.
That's the difference between "run this SQL" and "give me revenue by region using the metric we all signed off on." The second one is the only version where I'm inclined to trust the number that comes back.
get_chart_sql
Transparency is a governance requirement, not a nice-to-have. This tool hands an agent the underlying SQL and applied filters for any chart. So when someone asks how a number was calculated, and someone always asks, the AI can show them the exact query, the exact filters, and the exact data source. Most competitors just don't expose that through MCP.
What Preset MCP looks like in practice
Here's a workflow you can run today, with Preset's MCP server and Claude:
- An analyst opens Claude and connects to Preset over MCP.
- They ask: "What tables do we have related to customer churn?"
- Claude uses Preset's discovery tools to turn up the relevant datasets, schema and metric definitions included.
- They refine: "Build me a query for monthly churn rate by customer segment over the last 12 months, using our standard churn metric."
- Claude runs it through the semantic layer with
query_dataset, honoring the team's metric definitions. - They eyeball the results, tweak the segmentation, and say: "Save this as a virtual dataset called 'Monthly Churn by Segment' so the customer success team can use it."
- Claude calls
create_virtual_dataset, and now there's a governed dataset sitting in Preset, queryable by anyone with the right permissions.
Step 7 is the one no other BI vendor supports through MCP. The dataset doesn't vanish when the chat window closes. It becomes part of the organization's data layer, and the next person, human or agent, starts from it instead of from scratch.
Open source underneath
Preset's MCP server is built on SIP-187, an open-source specification we contributed to Apache Superset. Any MCP client that works with open-source Superset works with Preset, and the other way around.
This is the part competitors can't easily copy. Tableau's MCP lives in the Salesforce ecosystem. Looker's routes through Google Cloud. Power BI's is a resident of the Microsoft Fabric universe. Preset's works with any AI client, on any cloud, against any of our 75+ supported databases.
If you're running multi-cloud, or just trying to evaluate AI tools across your stack without swearing loyalty to one model vendor or cloud provider, that neutrality isn't a luxury. Your BI tool's AI shouldn't be the thing that locks you in.
Security isn't an afterthought
Every MCP tool call goes through the same security stack as the Preset application itself: OAuth 2.0 with PKCE, role-based access control, row-level security, and full audit logging. An agent connected over MCP sees exactly what its authenticated user is allowed to see, and nothing else.
The server inherits Preset's seven-layer middleware architecture, the same governance behind SOC 2 compliance, HIPAA-eligible deployments, and enterprise-grade access control. When an agent creates a virtual dataset, that dataset picks up the workspace's security model automatically. There's no writing your way around permissions.
The bigger picture
The whole BI market is circling one question right now: what role does your analytics platform play in an AI-native workflow?
Some vendors want to be the destination, where the AI features live inside their product UI. Others want to be the source, a governed data layer that any AI tool can point at. We dug into that framing in AI in BI: the path to full self-driving analytics.
My honest answer is: both. Preset Chatbot gives teams a conversational experience inside the product, grounded in the semantic layer. Preset MCP makes the platform composable, an open building block that agents, developer tools, and custom applications can plug into and build on.
Twenty-five tools that read, query, discover, and, the part that matters, create. On an open protocol, with an open-source foundation and enterprise security underneath. That's a longer answer than "we added a chatbot." That's rather the point.
Want to see it in action? Read the Apache Superset MCP technical deep dive, see the full Preset MCP announcement, or go under the hood in the Preset Chatbot technical deep dive.
