Query Streams MCP

Let your AI think.
We’ll handle the heavy data work.

Find the right tables. Turn large results into compact findings. Keep the data ready for the next question.

Query Streams gives your AI client a partner beside your databases: the Network Agent. Less data to read and move. Fewer repeated requests. More room to work through your question.

Start with Free See it in action

Table discovery and read-only queries on Free.
Profiling and follow-up analysis from Personal Core.

A Network Agent works beside two databases inside a building, sending a compact findings card to an AI client outside. — **The groundwork happens beside your data.**Your AI client uses the findings to reason, explain and follow up.

Why Query Streams MCP?

Because getting a useful answer takes more than running a query. Your AI client needs to find the data, make sense of it and ask the next question.

Find the right place to start.

A short question finds relevant tables by meaning, even among thousands of coded names. Your AI client reads the schema it needs and gets on with the query.

Less searching. Fewer guesses.

Spend tokens on useful context.

The Network Agent profiles and analyses large results beside your database. Your AI client receives compact findings it can use to decide what to explore next.

Less raw data to read and transfer.

Ask the next question without starting over.

Filter, group, compare or sample a held result. Follow-up tools reuse the data already on the Network Agent, without another database query.

Fewer round trips. Less repeated work.

Work with private data on your terms.

The Network Agent connects out from inside your network. Database credentials stay there, and queries pass a read-only check before they reach your data.

No inbound database ports or VPN.

Table discovery is included on Free

Start with the right tables.

Your AI client asks for what it needs in plain words, in any language. Query Streams returns a short, ranked list of matching tables, with a reason for each match. Then your AI client reads the columns it needs and writes the query.

Ask in your language. Find tables by meaning.Search names, columns and descriptions across a captured database. Semantic Intelligence can add meaning to abbreviated names, and your team’s table choices help improve future searches.
Works across the whole captured database.New captures become searchable within minutes. Table and description changes are reflected automatically. Finding a table does not query your database.

Measured in development tests on November 6, 2026: about 60 ms median per search; 71 of 72 questions across 12 languages found the right table in the top three. Results vary by database and question.

Your AI client asks

Find vehicle auction history, then count the records.

A ranked shortlist, with reasons qs_find_tables

auc_uaxmtn.conlibv_pfarbhist1,048 rows · Match in description
auc_uaxmtn.conlibv_pfvehcond2,258,576 rows · Name match: vehicle
Read as: vehicle condition

Next: read the matching table's schemaqs_get_table_schema database: asi table: auc_uaxmtn.conlibv_pfarbhist

Then: run the read-only querySELECT COUNT(*) AS history_rows FROM auc_uaxmtn.conlibv_pfarbhist;

Illustrative auction-database exchange using table names and row counts from development testing. Reply shortened for readability.

See what your AI can do with a head start.

Once the right data is found, the Network Agent can profile it, calculate over it and keep it ready for follow-up questions. Choose an example to see how a large result becomes something your AI client can explain.

Your AI client or Nova

Nova

Uses the model and decides what to ask.

Network Agent

Profiles and analyses data beside your databases.

Profile invoices, then prioritise collectionsSaved query: Invoice analysis

Illustrative analysis
Personal Core and up

Understand the invoice result
Profile the data before reading the rows.
Nova or your AI client needs to understand a large invoice result. It asks for a local profile, rather than reading batches of invoices to discover the columns and values.
The AI starts with a compact map of the data.
Nova or your AI client
What is in this invoice result, and which fields can help us chase overdue balances?
Understand the result before asking for the detail.
Saved query: Invoice analysis
Optimised tool: result profiling
Profile a large invoice result.
The agent scans the result beside your database and returns column statistics, ranges and common values. The AI can see how to analyse the invoices without processing every row itself.
Less raw data in the model context means fewer tokens to process and less data to transfer over the internet.
Network Agent, inside your network
Invoice analysisWorks withSQL Server
Profile a query resultqs_run_saved_query / mode: profile
184,902rows profiled
6columns
Selected column findings
Invoice ID 0 missing values
Amount 1,140.00 to 28,900.00
Status Open, Overdue, Paid
Region North, South, East, West
Complete result held locallyThe AI receives the profile. The agent keeps the rows ready for the next calculation.
Follow-up tool: query the held result
Ask for the totals that matter.
The profile gives the AI the fields it needs. It asks the agent to filter overdue invoices and group their balances by region, using the result already held locally.
No second database query in this example.
Nova or your AI client follows up
Group the overdue balances by region. Where should we start?
Use the held resultFilter overdue invoices and calculate SUM(amount) by region on the held result.
The agent reuses the same result beside your data.
Four totals, ready to explain
Put the largest balances first.
The agent hands back four regional totals. The AI can focus on collection priorities instead of repeatedly requesting, reading and adding up invoice rows.
184,902 rows stay local for this analysis. Four grouped rows answer the follow-up.
Network Agent, inside your network
Group overdue balances locallyqs_result / op: query
Region Overdue
South 28,900.00
North 17,320.80
East 9,765.25
West 5,493.95
Total 61,480.00
184,902 rows held locally. This follow-up returns 4 summary rows to the AI.
Your AI can now explain
South and North account for 75.2% of the overdue amount. Start collections there, then work through the smaller regional balances.

Selected column findings
Invoice ID	0 missing values
Amount	1,140.00 to 28,900.00
Status	Open, Overdue, Paid
Region	North, South, East, West

Region	Overdue
South	28,900.00
North	17,320.80
East	9,765.25
West	5,493.95
Total	61,480.00

Segment stock, then isolate replenishment gapsSaved query: Stock cover and delivery

Illustrative analysis
Business Core and up

Look across the whole catalogue
Find the stock that needs attention.
The AI asks for natural groups based on stock cover and delivery time. It does not need to pull 1,284 product records into its context to work out which ones behave alike.
The held result already includes stock cover and expected delivery intervals.
Nova or your AI client
Which stock lines behave alike, and which group is exposed before the next delivery?
Find the unusual group across the catalogue, then identify its products.
Saved query: Stock cover and delivery
Optimised tool: held-result segmentation
Find groups with similar stock cover.
The agent groups the numeric fields locally and returns the size and average values of each segment. One small group stands out: its average delivery arrives after its stock runs out.
Three segment summaries give the AI a focused place to investigate.
Network Agent, inside your network
Stock cover and deliveryWorks withPostgreSQL
Find natural groups in a held resultqs_result / op: segments
1,284 stock lines become three segment summaries.
- Short cover2 products
  Average cover 2.5 days
  Average delivery 4.5 days
- Steady cover882 products
  Average cover 18 days
  Average delivery 5 days
- Long cover400 products
  Average cover 45 days
  Average delivery 4 days
Segment names explain the illustrative numeric groups.
Follow-up tool: filter the held result
Zoom in on the exposed products.
The segment values guide a precise follow-up. The AI asks the agent to return products whose stock cover is shorter than their delivery interval, using the same held rows.
No second database query in this example.
Nova or your AI client follows up
Which products have less stock cover than their delivery time?
Use the held resultRead the held rows where stock_cover_days is less than delivery_days.
The agent reuses the same result beside your data.
Two products, two clear gaps
Close the replenishment gaps.
The agent returns the two matching products and their intervals. The AI can name the affected lines and explain the two-day exposure without sorting through the whole catalogue.
The agent does the grouping and filtering; the AI explains the action.
Network Agent, inside your network
Filter and calculate over held stock rowsqs_result / op: query
Stock coverGap until delivery
Packing boxes2-day gap
2 days of stock. Delivery in 4 days.
Mailing bags2-day gap
3 days of stock. Delivery in 5 days.
1,284 rows held locally. Only the two exposed products are returned for this follow-up.
Your AI can now explain
The short-cover group contains two exposed products. Both run out two days before delivery. Bring those replenishments forward; the other groups have much more cover.

Detect a dispatch shift, then compare periodsSaved query: Daily dispatch performance

Illustrative analysis
Business Core and up

Investigate a performance change
Ask what changed over time.
The AI starts with a daily dispatch result containing dates, late counts, shipped counts, late rates and period labels. It asks the agent to find changes in the late-count series.
The held result has 90 daily observations, not a new stream of individual orders to read.
Nova or your AI client
Why have late dispatches stayed high? Is this a one-off spike or a change in the pattern?
Detect a change over time, then test a more useful measure.
Saved query: Daily dispatch performance
Optimised tool: time-series anomalies
Detect a sustained change.
The agent checks the ordered numeric series for spikes and level changes. This example highlights a move from about three late dispatches per day to about twelve, giving the AI a pattern to investigate.
The AI receives detector findings instead of working through the full series itself.
Network Agent, inside your network
Daily dispatch performanceWorks withMySQL
Detect spikes and level changesqs_result / op: anomalies
Earlier patternChange detected
Level change in the late-count seriesAbout 3 to about 12 late dispatches per day.
Selected days from a 90-day held result. The agent returns flagged points and a summary.
Follow-up tool: compare held groups
Check the rate, not just the count.
A higher count could reflect more orders. The AI asks the agent to compare daily late rates across the two periods, using the period labels already present in the held result.
No second database query in this example.
Nova or your AI client follows up
Has the late-dispatch rate risen too? Compare the earlier and recent periods.
Use the held resultCompare late_rate between the two period groups in the held result.
The agent reuses the same result beside your data.
A comparison with useful context
Measure the change between periods.
The agent returns summaries for the 62 earlier days and 28 recent days. The AI can explain how the proportion of late orders changed, rather than treating every rise in the count as the same problem.
Fewer repeated data requests leave the model more room for the explanation.
Network Agent, inside your network
Compare late rates between held groupsqs_result / op: compare
Mean daily late-dispatch rate
Earlier period62 days2.0%
Recent period28 days6.2%
+4.2 percentage pointsThe proportion of late orders increased too.
90 daily rows held locally. Two group summaries shown from the comparison result.
Your AI can now explain
The mean late rate rose from 2.0% to 6.2%, an increase of 4.2 percentage points. A larger share of orders is late, so investigate the dispatch process during the recent period as well as order volume.

Illustrative data and simplified workflow. No MCP request is sent. Profiling and held-result queries require Personal Core or above. Segmentation, anomaly detection and comparisons require Business Core or above, within your plan limits. Free includes standard read-only MCP queries.

The same Network Agent groundwork supports Nova, our built-in AI analyst.

The Network Agent advantage

The heavy work belongs beside your data.

Profiles, calculations and held results stay close to the database. Your AI client receives the findings it asks for, with less data to transfer and less repeated processing.

Standard MCP queries return query results. With the optional analysis tools, your AI client can request a compact profile or calculation instead of the full result.

Your database credentials stay inside.Your AI client signs in to Query Streams or uses a revocable MCP key. It never needs your database password.
One outbound, encrypted connection.Requests and results travel over a connection opened by the Network Agent. Your database needs no inbound port or VPN for Query Streams.
Read-only is enforced beside the data.The Network Agent checks every query and refuses statements that would change your data.

Bring the data your business actually runs on.

Use the same MCP connection across supported databases, files, logs and API-connected services. Saved queries and cross-source queries give your AI client more ways to work with the data you already have.

Explore supported sources

DatabasesSQL Server, PostgreSQL, MySQL, Oracle and more, on your network or in the cloud.
Files and logsWork with CSV, Excel, Parquet and log data through supported file connectors.
Connected servicesQuery service data and join it with your other sources, within your plan limits.

Start free. Add deeper analysis when you need it.

Free

Find the right tables and run read-only queries. Use saved queries and join up to two sources.

Table discovery in plain language
One API key, 30 calls per minute
Up to 1,000 rows per query

Free is a plan. It never expires.

Personal Core

Give your AI client compact findings and data it can reuse.

Size checks and column profiles
Compact result summaries
Held results for follow-up questions

Less data work in the AI conversation.

Business Core

Go deeper into quality, patterns and changes across your data.

Data-quality checks
Anomalies, segments and comparisons
Nova, the built-in AI analyst

More ways to investigate what matters.

Features are available within each plan’s allowances. Compare plans and limits

Connect your data once. Bring your AI client.

Install the Network Agent on Windows, Mac or Linux and connect your data sources. Then connect your AI client through SSO or a revocable MCP key.

Sign in with SSO

For AI clients that support sign-in

Add Query Streams in your AI app, sign in with your Query Streams account and approve access. No API key to copy.

Connect with an API key

For tools that support API-key connections

Create a revocable MCP key in the portal's connection wizard, then add the server address and key to your tool's settings.

Server settings for API-key clients

"querystreams": { "url": "https://mcp.querystreams.com/mcp", "headers": { "X-MCP-Key": "your key" } }

Both methods keep your database login on the Network Agent. Revoke access from the portal any time.

Step-by-step guide Find Query Streams in your AI client’s connector directory Or use Nova, the AI analyst built into Query Streams

Questions people ask first

Does it work on a very large database?

Yes. Your AI client asks for what it needs and gets matching tables ranked, with a reason for each match, whether the database has 40 tables or 20,000. It can find the right table without reading the whole list.

Does my AI need to know my table names?

No. Your AI client can ask in plain words, in any language. Table names, column names, descriptions and captured sample values help find matches. Semantic Intelligence can add meaning to abbreviated or coded names, but it is optional.

Does finding a table use my data allowance?

Only the small reply itself. Finding a table reads nothing from your database.

What happens right after I connect a big database?

It becomes searchable within minutes of schema capture, with nothing to configure. Until then, your AI client is told the index is still building and how far along it is, and it can list tables the usual way. Changes to tables and descriptions are reflected automatically.

Does finding a table send my schema to an AI vendor?

Finding a table runs entirely inside Query Streams; no third-party AI service performs the search. Your connected AI client receives the matching table names and reasons. If it asks for a table’s schema or query results, those replies are returned to that client too.

Do I need to open ports or run a VPN?

No. The agent runs inside your network and opens one outbound, encrypted connection to Query Streams. Requests and results travel both ways over it. Your AI client connects to Query Streams, never to your database, so there is no inbound port, no IP allowlist and no VPN. The agent runs on a laptop, a VM, on-prem hardware or any cloud that allows outbound HTTPS.

Can I use SSO or an API key?

Yes. If your AI client supports SSO, connect with your Query Streams account. For clients that use API-key authentication, create a revocable MCP key in the portal and add it with the server address. Your database login stays on the Network Agent with either method.

Does my AI client see my rows?

Standard MCP queries return query results to your AI client. Optional profiling and analysis tools send compact findings instead of the full result. Held results stay on your Network Agent, and your AI client receives only the rows or summaries it requests.

What is a held result?

A result your AI client asks the agent to keep for a while. Follow-up questions are answered from that copy, with no second database query: filter it, group it, sort it, page through it or take a sample. Query Streams' servers never store those rows. It is released when its hold time ends, when the assistant releases it, or when someone clears held results on the Data Agents page. Held results are on Personal Core and up.

What if the AI tries to write or delete data?

Every query passes a read-only check on the agent before it reaches your database. SELECT, WITH and EXPLAIN run; anything else is refused. The check runs inside your network, not in the cloud.

Which plans include it?

Table discovery and read-only MCP queries are included on every plan, including Free. Free never expires and includes one key, 30 calls a minute and up to 1,000 rows per query. Profiling and held-result analysis start on Personal Core; deeper analysis starts on Business Core. Nova, the AI analyst built into Query Streams, starts on Business Core. Compare plans.

Give your AI a head start on its next answer.

Find the right tables. Let the Network Agent prepare the data. Give your AI client more room to think.

Get the free plan

Let your AI think.We’ll handle the heavy data work.

Why Query Streams MCP?

Find the right place to start.

Spend tokens on useful context.

Ask the next question without starting over.

Work with private data on your terms.

Start with the right tables.

Profile the data before reading the rows.

Profile a large invoice result.

Ask for the totals that matter.

Put the largest balances first.

Find the stock that needs attention.

Find groups with similar stock cover.

Zoom in on the exposed products.

Close the replenishment gaps.

Ask what changed over time.

Detect a sustained change.

Check the rate, not just the count.

Measure the change between periods.