Hero Background
');background-size:60px 60px">

Your agents shouldn't be writing SQL against production

Give them a menu of analytics you already trust instead.

SparteraConnect turns the analytics your team has already written and tested into a managed MCP server your own AI agents can call. They pick from a described catalog and pass parameters. No schema exposure, no generated SQL, no thousand-row payloads coming back.

4 bytes
Typical answer size
0 rows
Leave your warehouse
60 sec
To deploy the server

Why Spartera Connect

Three ways AI talks to data. Only one gets it right.

Most ways of connecting an agent to data either hand it raw rows or let it invent the query. SparteraConnect does neither.

01 — Traditional REST API

Raw Data Dump

API returns raw JSON. AI must figure out the rest.

🤖

Agent calls GET /api/sales

Requests raw data from endpoint

📦

Gets 10,000 JSON rows back

Full dataset transferred to consumer

🎲

AI tries to compute trends itself

LLMs are notoriously bad at math on raw data

❌

Returns wrong insight confidently

Hallucinated aggregation, no verification

response // 10,000 rows of raw data
[{"id":1,"amt":847.20},
 {"id":2,"amt":1203.50},
 // ...9,998 more rows
]
// AI guesses: "Revenue up 15%"
// Actual: Revenue DOWN 3%
Accuracy Low — AI guesses
⚠️
Your data moves. Your insights are wrong. AI computes analytics from raw rows and gets them wrong with full confidence. Data physically leaves your infrastructure.

02 — NL2SQL MCP Server

AI-Generated Queries

AI writes SQL on the fly. No one verifies it.

🤖

Agent receives question in English

"What's driving revenue growth?"

⚡

LLM generates SQL on the fly

Wrong joins, bad filters, missed tables

🗄️

Runs unverified query on production

Your schema is exposed, no sandbox

❓

Returns result — maybe correct

No way to know without manually checking

AI-generated // AI writes this SQL:
SELECT region, SUM(revenue)
FROM sales
WHERE quarter = 'Q4'
// ⚠ Missed the refunds table
// ⚠ Wrong fiscal year mapping
// → Numbers look right, but aren't
Accuracy Unreliable — 60-80% on clean schemas
⚠️
Confident but wrong. Who's accountable? NL2SQL accuracy tops out at 60–80% on clean schemas. In production with messy real-world databases? Much worse. Wrong numbers, no audit trail.
✦ Recommended

03 — Spartera Connect

Verified Analytics

Data owner builds & verifies. AI picks from a curated menu.

✅

Data owner pre-builds analytics

Stats, trends, predictions — tested & verified

🤖

AI selects the right analytic

Chooses from curated, parameterized menu

🔒

Runs against your warehouse

Only the computed result is returned — no rows, no schema

📊

Returns verified, auditable insight

Correct every time, production-grade

response // Verified analytic response:
{"insight": "Revenue up 23% YoY",
 "drivers": ["Enterprise +45%",
             "West region +31%"],
 "analytic": "revenue-yoy-drivers",
 "verified_by": "data owner",
 "latency_ms": 312}
Accuracy Verified by the data owner
✅
Right answers. Every time. Auditable. Analytics written and tested by the people who know the schema, executed against your warehouse. The agent receives the answer, not your data.
REST API NL2SQL Spartera
What reaches the agent 🔴 Full dataset 🟡 Raw rows 🟢 Result only
Results verified? 🔴 No 🔴 No 🟢 By data owner
Schema exposed? 🟡 Partially 🔴 Fully 🟢 Never
Accountability? 🔴 None 🔴 None 🟢 Full audit trail
Works at scale? 🟡 Data transfer bottleneck 🟡 Query complexity limit 🟢 Pre-optimized analytics
✅ Start free

Verified queries · No rows or schema to the agent · ~30 supported engines

The same question, two paths

Watch what comes back

Both paths answer the question. Only one of them keeps your rows, your schema and your context window out of it.

The question

“What percentage of 4th down attempts resulted in touchdowns last season?”

Path A — NL2SQL

Writes the query at runtime

Ready
Round trips 0
Rows returned 0
Into agent context 0tok

8,412 rows returned for the model to reason over

  1. Inspect the schema

    Reads tables and columns so the model can guess where the data lives

    SELECT table_name, column_name
FROM information_schema.columns

  2. Generate SQL from the question

    The model writes a query nobody has reviewed

  3. Run it against production

    Counts every 4th down play, including ones wiped out by penalty

    SELECT COUNT(*) FILTER (WHERE is_touchdown)
     , COUNT(*)
FROM plays
WHERE down = 4

  4. Return 8,412 rows

    The result set travels back into the model's context window

  5. Model does the arithmetic

    A language model computing a percentage over rows it never validated

Returns

28.7% of 4th down attempts resulted in touchdowns.

Wrong, and it won't say so. The generated query counted plays negated by penalty and scored two-point conversions as touchdowns. Plausible, confident, and 3.7 points off.

Path B — SparteraConnect

Calls a query that was already verified

Ready
Round trips 0
Rows returned 0
Into agent context 0tok

4 bytes returned — the measured median for a calculation

  1. search_metrics

    Reads the described catalog and matches the question to one analytic

    search_metrics(query: "4th down touchdown rate")
→ 4th-down-td-pct

  2. query_metric

    The verified query runs against your warehouse; the value comes back

    query_metric(
  asset_id: "4th-down-td-pct"
)

Returns

25.02% of 4th down attempts resulted in touchdowns.

Correct and reproducible. The query was written and tested by the person who knows the schema — including that penalty-negated plays and two-point conversions don't count.

  • analytic · 4th-down-td-pct
  • verified · tested before publish
  • payload · 4 bytes
  • schema · never left your warehouse

Payload returned

8,412 rows → 4 bytes

Median calculation response, measured across production calls.

What the agent sees

Schema → nothing

No tables, no columns, no credentials cross the boundary.

The answer

Right vs. plausible

Both sound confident. Only one traces to a query someone tested.

The 4-byte figure is real: it's the median response size for CALCULATION analytics measured across production calls. The NL2SQL side is an illustrative 8,400-row result set, and the accuracy point comes from published text-to-SQL research where accuracy tops out around 60–80% on clean schemas. Step timings are illustrative and not a performance claim — real query time depends on your warehouse and how complex the analytic is.

Interactive demo

Ask the same question both ways

Flip the switch in the header to change how the agent gets its answer. Same question, same data — different path, different result.

Spartera

Spartera IntelliChat

Analytics assistant

Agent reads the described menu of analytics and calls a verified one with parameters.

Try an example:

Analytics → MCP server Knowledge → RAG Concepts → LLM

Scripted demo using real analytics from the Spartera catalog. Token counts are illustrative. Deploy your own →

What the demo shows

Three exchanges from the demo above, written out.

“What percentage of 4th down attempts result in touchdowns?” Same question, two answers

NL2SQL

28.7%

The model wrote a query at runtime, pulled 8,412 rows back into its context window and did the arithmetic itself. The query counted plays negated by penalty and scored two-point conversions as touchdowns. The answer is wrong and nothing in the response says so.

8,412 rows returned · 5 round trips · schema exposed

Verified analytics

25.02%

The agent called search_metrics, matched the described analytic 4th-down-td-pct, and ran it with query_metric. The query was written and tested by the person who knows the schema. The answer comes back as a 4-byte value — the measured median for a calculation, with no rows to reason over.

4-byte payload · 2 round trips · schema never exposed

“Who scored the most touchdowns last year?” Ambiguous question

NL2SQL

Picks one interpretation and answers

The question doesn't say passing, rushing or receiving. The model resolves the ambiguity silently, writes a query for whichever it assumed, and returns a confident single name. You have no way to know a choice was made on your behalf.

Verified analytics

Asks which one you meant

Three analytics in the catalog match “touchdowns” — passing, rushing and receiving. Because each is described, the agent can see the ambiguity and ask rather than guess. You pick, then it runs that one.

“How many penalty yards did the Bills have?” Where NL2SQL wins

NL2SQL

Answers it

No one anticipated this question, and NL2SQL doesn't need them to have. It writes a query against whatever tables exist and returns a number. This is the real advantage of generating SQL at runtime: reach. Whether the number is right is a separate question, and one you'd have to check yourself.

Verified analytics

Says it doesn't have one

Nothing in the catalog covers penalty yards, so the agent says so and offers to queue it for the data team rather than improvising. A known gap beats a confident guess — but this is the trade, and it's why the two approaches belong together: explore with NL2SQL, promote what matters into verified analytics.

Four steps, and you own every one of them

Self-serve from signup to a running server

1
🔌

Connect a data source

Point Spartera at the warehouse or database the analytics should run against. Credentials stay on your side; we never copy your data.

Varies with your environment
  • ✓ Warehouses: BigQuery, Snowflake, Databricks SQL, Redshift, Synapse, Teradata
  • ✓ Databases: PostgreSQL, MySQL, SQL Server, MariaDB, Aurora, Cloud SQL, Supabase
  • ✓ Also: ClickHouse, Dremio, Apache Hive, Apache Impala, S3, SageMaker, Databricks Model Serving
  • ✓ Structured data and model endpoints — not documents or PDFs
2
✅

Write and test your analytics

Author the queries you want agents to be able to call, and describe what each one answers. This is the part that makes the difference — the description is how an agent knows which analytic to pick.

Depends how many you build
  • ✓ Calculations return a value; visualizations return a rendered chart
  • ✓ Parameters are declared and bound at runtime, never concatenated
  • ✓ Test each one before it becomes callable
  • ✓ Customers have built 18,751 analytics on the platform so far
3
🚀

Deploy the MCP server

One button. We provision an isolated Cloud Run service, wire up authentication and rate limiting, and hand you a connection string.

About 60 seconds
  • ✓ Dedicated server per company — no shared tenancy
  • ✓ Bearer-token authentication and per-server rate limiting
  • ✓ Encrypted in transit and at rest on Google Cloud
  • ✓ Latency and usage recorded on every call
4
🤖

Point your agents at it

The server exposes four tools. Your agent lists what's available, reads the definitions, and runs the one it needs.

Immediate
  • ✓ list_available_metrics — everything the agent is allowed to call
  • ✓ search_metrics — find the right analytic from a natural-language question
  • ✓ get_metric_definition — read parameters and what the analytic returns
  • ✓ query_metric — execute it with bound parameters
  • ✓ Works today with Claude Desktop over stdio, or any client you configure yourself

See It In Action

❌

Without SparteraConnect

Customer:
What's driving our revenue growth this quarter?
AI:
I don't have access to your revenue data. Please check your BI dashboard or contact your analytics team.
❌ Executive frustrated, waits 3 days for analyst report
✅

With SparteraConnect

Customer:
What's driving our revenue growth this quarter?
AI:
Your Q4 revenue is up 23% YoY, driven primarily by the Enterprise segment (+45%) and West region (+31%). SaaS subscriptions grew 18% while professional services increased 12%. The growth accelerated in November with the product launch, contributing an additional $2.1M. Would you like me to break down by customer cohort or compare to industry benchmarks?
✅ Executive gets insights in 3 seconds, makes informed decisions
🚀 Start free

30 analytics calls a month, no card required

How Is This Different From RAG?

They're complementary—RAG for documents, SparteraConnect for analytics

You've probably heard of RAG (Retrieval Augmented Generation)—it's great for searching through documents, PDFs, and knowledge bases. SparteraConnect does something completely different: it pulls analytics and insights from structured databases. Think of them as partners, not competitors.

RAG
Document Search
SparteraConnect
Database Analytics
Best For
Unstructured documents & text
"What's our return policy?" (from PDF)
Structured data & analytics
"What's our return rate this quarter?" (from database)
Data Source
PDFs, docs, articles, knowledge bases
Company wiki, product manuals
SQL databases, data warehouses, ML models
PostgreSQL, Snowflake, BigQuery
Query Type
Semantic search & retrieval
Find relevant paragraphs
SQL queries & calculations
Calculate metrics, aggregate data
Output
Text passages from documents
"According to the handbook..."
Analytics, metrics, insights
"Revenue grew 23% vs last quarter"

Use Them Together for Complete AI

The best AI assistants use both. RAG answers questions from your documentation. SparteraConnect answers questions from your data. Together, your AI can handle policies AND metrics, documentation AND analytics.

Example Scenario

E-commerce business intelligence AI

📚
RAG Handles:
  • • "What's your shipping policy?" → Searches company docs
  • • "How do I return an item?" → Retrieves return process from PDF
  • • "What payment methods do you accept?" → Finds info in knowledge base
📊
SparteraConnect Handles:
  • • "What's our customer churn rate by cohort?" → Analyzes data warehouse
  • • "Which products are trending up this quarter?" → Computes sales trends from database
  • • "What customer segments have highest LTV?" → Runs ML model predictions
✅
Result:

One AI that answers both documentation questions AND analytics questions

What happens when you point an agent at your warehouse

Every approach that lets an agent reach your data has a failure mode. Most of them are quiet.

🎲

Confidently wrong numbers

An agent that writes SQL at runtime will answer anything, including the questions it gets wrong. Published text-to-SQL accuracy tops out around 60–80% on clean schemas and falls further on real ones.

🔓

Your schema in someone's context window

Generating SQL means introspecting tables and columns first. Your data model ends up inside a model prompt, and so do the rows it reads back.

📦

Payloads that swamp the agent

Thousands of rows come back so the model can do arithmetic on them. It burns context, costs tokens on every call, and asks a language model to do a database's job.

⏳

Or you build it yourself

A production MCP server with auth, rate limiting and audit logging is a real engineering project — and then you maintain it.

What SparteraConnect does instead

The queries are written once, tested, and then called by name.

✅

Verified, not generated

Your team writes and tests the query. The agent binds parameters to it.

Same query every time, by the people who know the schema
🔒

Nothing leaves but the answer

The query runs against your warehouse; only the computed result returns.

No raw rows, no schema, no credentials to the agent
🧮

Answers, not rowsets

A calculation comes back as a value, not a result set to reason over.

4-byte median payload, measured across production calls
🔌

Across your whole stack

One server over BigQuery, Snowflake, Databricks, Redshift and more.

~30 supported engines, not one vendor's warehouse
💬 Put it in front of your agents

Start free, or talk to us first

What it costs

One subscription per server, plus what your agents actually call

Start free with 30 analytics calls a month. Deploying an MCP server adds a subscription.

BASE FEE

MCP server subscription

$100 /month per server

A dedicated, managed MCP server for your agents

  • ✓ Isolated Cloud Run service — not shared tenancy
  • ✓ Bearer-token auth and per-server rate limiting
  • ✓ Encrypted in transit and at rest on Google Cloud
  • ✓ Latency and usage telemetry on every call
  • ✓ Automatic scaling and patching
  • ✓ Support direct from the team that built it
Get Started
Metered

Plus calls

$0.50 per analytics call at list price

Volume brings that down substantially — talk to us if you're planning high call counts.

🆓 First 30 calls a month are free
📉 Volume pricing well below $0.50
🎯 No minimum, no contract, cancel anytime

💡 Scales with your usage automatically

See detailed API pricing →

Compared with building it

❌ Build it yourself:

  • • Engineering time to build the server, auth and rate limiting
  • • Published estimates for a production MCP server start around $100K
  • • Ongoing hosting and on-call
  • • You maintain it as the MCP spec changes

✓ SparteraConnect:

  • • About 60 seconds to deploy once analytics are defined
  • • $100/month per server, plus metered calls
  • • Infrastructure, auth and scaling included
  • • We track the spec so you don't

Frequently Asked Questions

What's free, exactly?

30 analytics calls a month, so you can build and test analytics as APIs before committing. Deploying an MCP server starts the $100/month subscription.

How are calls priced?

$0.50 per call at list. Volume discounts are significant — if you're planning hundreds of thousands of calls, the effective rate is a fraction of that.

Which AI clients can connect?

Claude Desktop connects over stdio today, and any agent you configure yourself can connect over HTTP. Remote connector directories require OAuth 2.1, which is on the roadmap.

Can I cancel anytime?

Yes. No contracts and no penalties — cancel the server subscription whenever.

Stay Ahead of the Analytics Revolution

Get insights on data commerce, AI grounding, and the future of proprietary data

We respect your privacy. Unsubscribe at any time.

Stop letting agents guess at your numbers

Start free

Sign up, connect a source, build a few analytics, and deploy a server when you're ready. No call required.

🆓
30 free calls a month
🔒
Rows and schema stay put
✅
Queries you tested yourself
🔌
~30 supported engines
🚀 Start free 💬 Talk to us

30 analytics calls a month are free. Deploying an MCP server starts a $100/month subscription per server, plus metered calls.