What changed, and when: “this metric broke on the 14th”
Spikes and level changes found in any series your AI client is holding, with no setup and no training.
What your Network Agent does for you
Your AI client can point at the day a metric broke. Your Network Agent scans one number column of a held result in time order and flags the points that stand out, each with its time, value, kind and score, plus a one-sentence summary. The series itself never leaves your network.
| Kind | What it looks like |
|---|---|
| Spike | One point far above or below its neighbours: a day of double orders, or a sudden dip. |
| Level change | A step that stays: returns went from 2% to 5% and stayed there. |
| Situatie | Before | With anomaly detection |
|---|---|---|
| “Is anything odd in daily sales?” | The client reads a chart’s worth of rows and guesses. | The unusual days, in time order, with a score. |
| “When did returns start rising?” | Guessing with date filters, one query at a time. | The date of the level change. |
| A year of daily totals | Too many rows to read. | A handful of flagged points and a sentence. |
The plan it needs
| Plan | This capability | Opmerkingen |
|---|---|---|
| Gratis | Not on this plan | Works as it always has, with standard results. |
| Personal Core, Personal Plus | Not on this plan | Context management and Result analysis are included; Deeper analysis is not. |
| Business Core, Business Plus, Business Max | Inbegrepen | Part of Deeper analysis. |
| Enterprise Core, Plus, Max | Inbegrepen | Part of Deeper analysis. |
Why it is on paid plans: these are analyses a chat client cannot do by reading rows. Your Network Agent runs them over the result it is holding, inside your network, and gives your AI client the groundwork to reason about: counts, scores, tests and examples.
The Free plan works as it always has. Paid plans add the Analytic toolset. Free is a plan you keep: MCP, the REST API, federated queries and alerts, with the same results as before. Plan names and prices are on the prijspagina.
How to ask
You do not call tools yourself. Ask in plain words and your AI client picks the tool. Three asks that bring this capability in:
Claude
ChatGPT
Cursor Claude Code+10 moreAlleen-lezenHold daily order totals for the last year.
Find anything unusual in that series.
When did the return rate start climbing?
For people who read tool lists: qs_result_anomalies with a held result id, time_col, value_col and an optional sensitivity from 0.1 to 0.9. See it on the Tools tab.
What comes back
Claude ChatGPT Cursor Claude Code+10 moreAlleen-lezenAnything unusual in daily orders this year?
3 anomalies in daily orders (2 spikes, 1 level change):
| datum | bestellingen | kind | score |
|---|---|---|---|
| 2026-03-14 | 1,940 | spike | 5.2 |
| 2026-06-02 | 310 | spike | 4.6 |
| 2026-08-19 | 1,420 | change | 3.9 |
Since 19 August orders have stayed higher. Want them split by channel from that date?
- points
- Each flagged point with its time, value, kind (spike or change) and score, in time order.
- spikes, changes
- How many of each were shown; if the list was cut at its cap, the full count is given.
- gescand
- How many points were examined. Rows that share one time are folded into one point first, and the answer says so.
- summary
- One sentence your AI client can quote.
What it does not help with
- It needs a time-ordered result: a date, timestamp or sequence column and a number column.
- It finds unusual points; it does not explain them. Your AI client then looks for causes with follow-up questions on the held result.
- Very short series give weak results, and a series needs at least a few points.
- It does not forecast.
Grenzen
- Sensitivity runs from 0.1 to 0.9; higher flags more. The default is 0.5 unless your organization changed it.
- The analysis stops after 30 seconds; narrow the original query and hold it again.
- The result must still be held: 60 minutes on Business plans, up to 8 hours on Enterprise.
Instellingen
On the Deeper analysis card on the Context tab: Anomaly sensitivity (0.1 to 0.9, default 0.5; higher flags smaller spikes and shifts) and the Analysis tools list, where switching Anomalies off refuses it for every key. A key’s own Analysis tools setting can narrow it further.
When your AI client says no
When the server refuses, your AI client receives a code and one sentence, and usually relays it. Quoted sentences are the server’s own words; the parts in italics stand for the ids a real message names.
| Code | What the server says | Wat te doen |
|---|---|---|
MCP_FEATURE_NOT_INCLUDED | “Your plan does not include advanced analytics; anomalies, segments, comparisons and result diffs are available on Business Core.” | Ask your AI client to work from a profile, a follow-up query or rows instead, or move up to Business Core. |
MCP_FEATURE_NOT_INCLUDED | “qs_result_anomalies is switched off for this organization; a workspace admin can turn it on.” | An admin turns it back on under Analysis tools on the Deeper analysis card of the Context tab. |
MCP_FEATURE_NOT_INCLUDED | “qs_result_anomalies is switched off for this key; the key’s owner can allow it on the key’s settings.” | The key’s owner edits the key on the MCP page and adds it under Analysis tools. |
MCP_AGENT_ANALYSIS_DISABLED | “analysis is switched off for the connector behind result {result id}; a workspace admin can allow it on the connector’s AI results settings.” | An admin opens the connector under Gegevens → Gegevensagenten and allows analysis in its AI results settings. |
MCP_RESULT_EXPIRED | “result w_3f9a0c12d47be801 is not held (it expired, was released, or is unknown here); call qs_result_windows for the current list, or run the query again with hold.” | Ask your AI client to run the query again. That is a new query on your database. |
MCP_WINDOW_SQL_REJECTED | “the analysis could not be expressed over the held result (a column named is not in the result, or its type does not fit the analysis).” | Name a column the held result has, of the right kind for the analysis. |
MCP_WINDOW_QUERY_TIMEOUT | “the analysis over the held result did not finish within 30 seconds and was stopped; narrow the result (a WHERE clause on the original query, fewer columns) and hold it again.” | Hold a narrower result and ask again. |
MCP_AGENT_NIET_VERBONDEN | The Network Agent that owns the connector is not connected right now. | Open Gegevens → Gegevensagenten and start the agent on its computer. |
Voorbeelden
“Why was March so good?”
Your AI client finds one spike on the 14th, then asks a follow-up on the held result and sees a single customer’s large order that day.
“Did the new returns policy change anything?”
A level change on the day the policy started: the daily return rate rose and stayed higher.
Related
Facts source
Facts on this page: the Query Streams MCP server and Network Agent as of 19 October 2026, with the server’s own refusal sentences quoted word for word. Table and column names in the examples are made up.
Setup trouble?
- Agent shows Offline, or Test connection and Browse are greyed out? Check that the Agent is online → Is the Agent online?
English 
Spanish 
German 
French 
Portuguese 
Japanese 
Korean 
Chinese 
Italian 
Dutch 
Indonesian 
Polish