Agents and assistants

Give your agents a working model of your business.

RootCause runs a standard MCP server over your governed ontology and your causal models. Any MCP client queries metrics by name, gets answers traceable to the rows and the model version behind them, and can ask what would happen next quarter.

Try it in the sandbox Read the MCP docs

Assistantwarehouse · SQL

If we cut lead time by three days on the west coast, what happens to churn next quarter?

I compared eighteen months of order and churn history for west coast accounts, grouped by average lead time.

−11.2%

churn gap between the fastest and slowest lead time quartiles

Accounts in the fastest quartile average 4.6 day lead times and renew at a higher rate.

Answered a different question.This is a historical association between lead time and churn. No intervention was modeled, so the number does not say what cutting three days would do next quarter.

AssistantRootCause MCP

If we cut lead time by three days on the west coast, what happens to churn next quarter?

Cutting three days off west coast lead time moves next quarter churn by:

−6.8% to −9.4%

modeled change in churn_rate, 2026 Q4, 95% interval

Modeled churn_rate vs today95% interval

today+13 weeks

modeled churnintervallead time unchanged

The effect concentrates in accounts carrying high support load. The interval widens toward the end of the quarter because the change moves those accounts outside observed conditions. Assumes current product mix and support staffing.

12 sourceschurn_rate v7model v42.1M rows

The same question typed into both panels. Open the tool trace to see the arguments and what came back.

Raw data

A schema leaves the agent guessing.

Point an agent at a warehouse and it will write SQL. Often good SQL. It will also pick the wrong one of your four customer tables, quietly include test accounts, and answer with the sales team's definition of revenue when the CFO asked for finance's. All of it arrives in a paragraph that reads as authoritative.

Nothing in a schema records which table is canonical or what revenue means in this company. The agent has to guess. A capable model guesses plausibly, which takes longer to catch than a bad guess.

An ontology removes the guess. The agent asks for a named metric with named dimensions. The platform compiles the query, applies access rules and returns the number the business already agreed on.

Traceability

Every answer resolves to the rows that produced it.

Which sources, which metric version, which model version, how many rows, how fresh. Expand a lineage chip and the contributing records are there. An analyst can check it, an auditor can reconstruct it, and a CFO can put it in a board pack and survive the follow-up question.

Agent answer

West coast service changes explain 7.9% of the modeled churn movement next quarter.

Sources that contributed to this answer

Source	Rows	Last sync	Role
CRM.accounts	413,880	12 min ago	account grain, region
Billing.subscriptions	1,142,004	8 min ago	churn event, contract value
Support.tickets	221,470	19 min ago	support load index
Logistics.shipments	986,213	31 min ago	lead time days

Four of twelve sources shown, ordered by contribution. The client can list all twelve.

Speculative questions

Ask what would happen if you changed something.

Governed retrieval answers questions about the past. Planning meetings run on the other kind. What happens if we move this. Which lever is worth pulling. How much of last quarter was us and how much was the market. Those questions need a model of how the business responds to a change.

Question

What happens to on-time delivery if we add a second carrier across the midwest lanes?

Answer over RootCause MCP

Modeled improvement of 4 to 7 points overall, concentrated in two of six lanes. Freight cost rises 2 to 3 percent.

Assumes current volume mix. 95% interval on the overall figure.

Intervention

Two lanes carry most of the service gain.

todaygain from a second carrier

Read how the models are built

Standard MCP

Any MCP client connects with one config block.

RootCause runs the protocol as published, with no bespoke SDK and no proxy layer to maintain. Tools are named for what they do: list the governed metrics, resolve an entity, query a causal model, simulate an intervention. Access rules follow the user through whichever client they are sitting in.

ClaudeDesktop, web and Code

ChatGPTDesktop and web

GitHub CopilotVS Code and JetBrains

CursorEditor and agent mode

WindsurfEditor and Cascade

ZedEditor and agent panel

Your own agentsAny framework, any language

Client names and marks are the property of their owners and are shown to indicate protocol compatibility.

Governance

Access is enforced before the query runs.

Row and column level rules apply at compile time, per user, for agents exactly as for people. An agent working on behalf of a regional manager sees that manager's regions. A query that would breach a rule is refused rather than filtered after the fact.

Every agent query is logged with the identity it ran under, the metrics it touched, the model version it used and the answer it returned.

Sandbox

Connect a client to the live sandbox.

The sandbox runs on a public foundational ontology. Paste the config into your MCP client, ask questions, expand the lineage, run an intervention. None of your data is involved and no sales conversation is required.

Open the sandbox

mcp.json

1{
2  "mcpServers": {
3    "rootcause": {
4      "url": "https://sandbox.rootcause.ai/mcp"
5    }
6  }
7}

Start here instead

Automated Data Ontologies

What the agent is actually querying, and how it gets built.

Next on the ladder

Simulation Engine

The full range of speculative questions, with worked examples.

RootCause MCP

Point your agent at a real question.

Bring one question your current setup answers confidently and wrongly.

Try it in the sandbox