Analyse Your Data with Your AI Client via MCP

Analyse your data with your AI client

Your AI client asks. Its partner inside your network does the heavy lifting.

Your Network Agent now does the measuring, profiling, holding and analysing where your data lives. Your AI client walks into every question with the groundwork done and spends its attention on the thinking. And nothing dials in.

Connect: every plan Analytic toolset: paid plans Rows stay in your network Alleen-lezen, altijd

Loop je ergens vast? Vraag het ons gerust. Open Ondersteuning Ga naar Query Streams (menu rechtsboven → Ondersteuning) en stuur ons een bericht. Echte mensen lezen het, en "Ik weet niet wat er op dit scherm wordt gevraagd" is een volkomen terechte vraag.

Your AI client can only read so much at once

Every AI client, whichever one you use, works inside a fixed context. Everything it reads in a conversation, your question, its answer, and every result a tool hands it, has to fit. The context is large for text, but data is dense: a single wide row from a business database can cost as much as a paragraph. In practice an AI client can take in a few hundred to a thousand rows of a real table before the rest is cut off, and when a result is too large most clients truncate it silently, drop it, or stop and ask you to narrow the question.

That is why “connect my AI to my database” demos work on a 50-row table and fall apart on the first real one. Ask for this year’s orders and the client gets the first thousand, says something confident about them, and never sees the rest. Ask it what tables exist in your ERP and the answer alone is bigger than it can read.

Nothing about this is a flaw in your AI client or in your data. No amount of prompting changes how much a client can read. The way around it is to stop handing the client raw rows and have the heavy, repetitive, context-eating work done before the client ever reads a thing, so it walks into every question prepared and spends its attention on the thinking.

What a typical database connector does about it

Very little. Most database connectors for AI hand the model a fixed set of record operations, or raw rows, and need a reachable endpoint into your network. Every follow-up is another database call. On the Free plan, Query Streams works this way too: standard results, read-only, through your Network Agent.

What Query Streams does about it on paid plans

Your Network Agent sits inside your network next to the database. It sees the whole result before any of it is sent, so it can measure it, profile it, hold it and analyse it there, and give your AI client what fits and what matters instead of the first thousand rows. Together these are the Analytic toolset.

Context management (Personal Core and up)

Schema that fits. Instead of every table and column at once, a short list of business areas, then the tables in one area, then full detail only for the tables your AI client names.
Results shaped to fit. Long text trimmed with a marker, binary values summarised, empty columns dropped, and only the columns the client asks for. Same rows, a fraction of the space.
A size budget, never a silent cut. Every result reports its size. The client can set a budget, whole rows are kept until it is spent, and what did not fit is reported with a link to the complete result.
Fits the client you use. The default result size is tuned to the client calling, so it works on the first try without anyone touching a setting.
Ask for the shape first. Columns, row estimate and size per row, with no rows sent, so the client knows it is looking at millions of rows before it asks for them.

Result analysis (Personal Core and up)

Profile a result of any size in one call. Your agent reads the whole result and sends back counts, ranges, averages, most common values, empty rates and a small sample. The rows never leave your network.
Follow-up questions without re-running. Your agent holds the complete result for a while. “Now break that down by region”, “show me the next page”, “give me a random sample of the cancelled ones” are answered from the held result, with no second query on your database.
Only what changed. Ask when a table was last updated, or re-run a question and get only the rows that changed since last time.

Deeper analysis (Business Core and up), computed by your agent over a held result, each answer with a one-sentence summary

Data-quality report. Empty and blank rates, format consistency, duplicate keys, outliers and a leading-digit check on amounts, ranked by severity.
Anomalies. Spikes and level changes in a value over time. “This metric broke on the 14th.”
Segments and comparison. Natural groups of similar rows, and whether two groups really differ, with the test, its p-value and the size of the effect.
Diff. What was added, removed, changed and unchanged between two runs, with example keys.

What this means in a conversation

Before: “Show me West region orders this year.” The client gets 1,000 of 200,000 rows, cuts most of those, and answers about a slice it did not choose.

After, on Personal Core: the client asks your agent for the shape, sees 200,000 rows, asks for a profile, and reasons from the whole result: the total, the refund rate, the top customers, the odd month. “Break the refunds down by customer” is answered from the held result a moment later. Your database ran one query.

Two halves: Connect and the Analytic toolset

Connect, on every plan

Every database, every AI client: Claude, ChatGPT, Cursor, Copilot and any other MCP client. Your AI client lists your connectors, reads the schema, and runs read-only queries through your Network Agent, with no database password in the AI and no firewall change.

The Analytic toolset, on paid plans

Your Network Agent does the groundwork a client cannot do by reading rows: schema by business area, results sized to fit, shapes, profiles, held results and freshness on Personal Core and up; data-quality reports, anomalies, segments, comparisons and diffs on Business Core and up.

The Free plan works as it always has. Paid plans add the Analytic toolset. The Free plan works as it always has: MCP, the REST API, federated queries and alerts, with the same results as before. Free is a plan you keep, not something that runs out.

Personal Core and up: Context management and Result analysis.
Business Core and up: Deeper analysis on top.
Enterprise: larger held results for longer, up to 4,096 MB per agent, held for up to 8 hours.

Why the Network Agent makes the difference

A typical database connector for AI

A fixed set of record operations, or raw rows. An endpoint the AI must be able to reach inside your network. Every follow-up is another database call, and the meaning of your columns is typed by hand.

Query Streams, with a Network Agent

One outbound connection and no open ports. The AI writes real SQL, and the agent reads the whole result next to your data and sends back what fits. Semantic Intelligence describes your data for you.

Faster because nothing crosses the wire that does not need to. Safer because the rows stay home.

Every capability, the plan that unlocks it, and how to ask

Capaciteit	Family	Plan	Ask it like this
Schema that fits	Context management	Personal Core and up	“What business areas does the ERP have?”
Token budget	Context management	Personal Core and up	“Keep the answer under 8,000 tokens.”
Column and cell shaping	Context management	Personal Core and up	“Only the date and status columns, first 100 characters of the notes.”
Client-aware defaults	Context management	Personal Core and up	Nothing to ask: the default fits your client.
Shape first	Context management	Personal Core and up	“How big is that table? No rows yet.”
Profile a result	Result analysis	Personal Core and up	“Profile this year’s orders.”
Follow-up questions on a held result	Result analysis	Personal Core and up	“Keep that result. Now break it down by customer.”
Data freshness and changed-since	Result analysis	Personal Core and up	“Anything new since this morning?”
Data-quality report	Deeper analysis	Business Core and up	“Is the customer table clean enough to use?”
Result diff	Deeper analysis	Business Core and up	“What changed in the price list since this morning?”
Anomalies and change detection	Deeper analysis	Business Core and up	“Anything unusual in daily orders this year?”
Segments and statistical comparison	Deeper analysis	Business Core and up	“Group our customers. Is the West really better?”
AI-ready Semantic Intelligence	Semantische intelligentie	Personal Core and up, with Semantic Intelligence run	“Profile the carrier table.”

Every tool stays listed for every key. On a plan that does not include a capability, the server answers with one sentence naming the plan that does, and your AI client carries on with what your plan has.

Measured on a real ERP database

A test database of 60 to 80 million rows per table.

Resultaat	What it was
Under a second	The shape of a 60-million-row table, with no rows sent
About 2,400 tokens	A profile of a million rows: ranges, top values, a fair sample
One call away	Long text trimmed to fit, with the full value one call away
One query	“Break it down by region” answered without touching your database

Where you set it up and tune it

Connect. On the MCP-pagina, Connect an AI client opens a short wizard: how your AI client will connect, which client, the setup for it, then a check that the connection works.
Context tab. Tune what your AI client receives: text limits, binary values, empty columns, number precision, schema view, result size and the Deeper analysis defaults, for your organization and for each key. Held results are listed there too, with a release button.
Tools tab. Every tool your AI client can call, what it returns, what is on your plan, and how much you use it.
Data Agents. Each agent’s Local analysis panel turns held results on or off, sets their budget and hold time, and has a Clear held results button.

Liever helemaal geen AI-client aansluiten? On Business Core and up, Nova is the AI analyst built into Query Streams. It uses the same Network Agent capabilities your AI client has, and it starts with more context: your Semantic Intelligence, your saved queries and your federated reports. Its results go to your screen as grids and charts, so it has no reading limit, and with Apply to Nova switched on it follows the same result rules you set for AI clients. Concierge-modus lets people who never write SQL ask in plain business language. With Aangeleerde kennis, you teach Nova a business rule once and it remembers it for the whole team; AI clients connected through MCP see the column descriptions Nova learns, but the business rules are applied only by Nova.

What leaves your network

The analysis runs on your Network Agent, inside your network. Only the answer and the rows a user asks to see are sent back. Read-only is enforced on the agent: only SELECT, WITH and EXPLAIN run. Every call is logged on the MCP Activiteit tab. These features need the latest Network Agent; agents update themselves. How Query Streams keeps your data secure.

Zit je nog steeds vast? Open Ondersteuning Ga naar Query Streams en laat ons weten in welke stap u zich bevindt en wat u ziet. Dan zorgen wij ervoor dat u verbonden wordt.

Facts source

Setup trouble?

Agent shows Offline, or Test connection and Browse are greyed out? Check that the Agent is online → Is the Agent online?