Working with AI agents
Once your agent is connected to gdotv's MCP server, you can ask it about your data in plain language. This page shows how to get good answers, and how agents work alongside you in gdotv: in your query tabs, on dashboards and on reports.
How it works
When you ask a question, the agent usually:
- finds the right connection and reads its schema, so its query uses your real labels, relationship types and properties;
- writes a query in the connection's language (Cypher, Gremlin, SPARQL, GQL, SQL…) and, for a new or complex query, checks it with gdotv and looks at the database's plan before running it;
- runs it through gdotv and reads a summary of the result;
- answers you, with a View in gdotv link that opens the complete result in gdotv.
Every query runs through gdotv with your access, and on every connection agents can only read unless you allow writes.
Start with the schema
Agents give better answers when they know what is in the graph. Start a conversation by asking about it:
What's in my Movies connection? Show me the schema.
Which relationship types connect people to movies, and what properties do they have?
How many nodes of each label are there?
The agent can also give you a link to the schema in gdotv's data model view.
Ask specific questions
The more context you give, the better the query. Say which connection, the time range, the filters and the format you want:
| Vague | Specific |
|---|---|
| Who are the top actors? | In Movies, which 10 actors appear in the most films released since 1990? Show a table. |
| Find fraud | In the Payments graph, find accounts that sent money to more than 5 new accounts in the last 7 days. |
| What's connected to Alice? | Show everyone within two hops of the person named Alice, and give me a link to see it as a graph. |
Ask follow-up questions to refine the result, as you would with a colleague: "only include films rated above 7", "now group it by year".
Expensive queries
Before running a new or complex query, agents can check it in one step: gdotv reads it for mistakes (syntax, and on gdotv Team and Enterprise unknown labels and properties) and asks the database for its plan, without running the query. The plan shows whether the query would scan every node or multiply two unrelated patterns, so the agent can fix it before it costs anything. Most RDF stores and SQL-based databases have no plan gdotv can get without running the query; there, the check covers the query text only.
When a query is slow, the agent can run it again with the database's profiler to see where the time went, on databases that support it (Neo4j, FalkorDB, Kuzu, Gremlin databases including Amazon Neptune).
This is slow. Profile it and tell me why.
Large results
Agents receive a summary of each result: the totals and a small sample (25 nodes, 25 relationships and 25 rows by default). gdotv keeps the full result, so the agent can page through it if it needs to. For large results, the best place to look at the data is gdotv itself: ask for the link.
Give me a link to see the whole result in gdotv.
Agents are told to count and aggregate rather than pull raw rows, and to start exploring with a LIMIT, which keeps answers quick.
Work on your open query tab
An agent can work directly on the query tab you have open in gdotv. It rewrites the query in the tab and runs it there, so you watch the text and the results change, and you can edit the query yourself at any point.
There are two ways to start:
- Ask your agent to "work on my current gdotv query tab". It can see which tabs you have open and which one is on screen.
- Click AI Agent in the query editor toolbar. gdotv shows a prompt that points your agent at this exact tab (or saved query, or dashboard panel). Add what you want it to do, then copy the prompt into your agent.
While the agent edits a tab or dashboard, gdotv shows a chip with its name, for example Claude Code editing. If you change the query while the agent is working, the agent sees your changes before it writes over them.
An agent can also put a query in your tab without running it, for example a write it wants you to review and run yourself.
Build dashboards
Ask your agent for a dashboard and it builds the whole thing: tabs, panels, queries, layout and parameters.
Build a dashboard for the Movies connection: releases per year as a line chart, the top 10 actors as a bar chart, and a table of the latest films. Let me filter by year.
The agent creates it as a private draft that only you can see, and gives you a link. Open it to watch the draft change as the agent works. The agent runs each panel to check that its result suits the panel type, for example that a bar chart has a label and a number to plot.
When you are happy, tell the agent to publish it, or click Publish in gdotv. Publishing makes it available to everyone who uses the connection. Agents also edit published dashboards through a draft: the published version stays as it is until the draft is published.
If you have unsaved changes on a dashboard when an agent publishes it, gdotv keeps your changes in a new draft rather than losing them.
Build reports
A report is a saved query with parameters, which the connection's users run from Saved Queries, filling in values such as a date range or a name. It is the best way to hand a question over for good: people who can't write queries can still run it.
Turn this into a report called "Top actors since a year", with the year as a parameter defaulting to 1990.
As with dashboards, the agent writes a private draft and gives you a link. You can run the draft in gdotv, then publish it when it is right. The agent sets each parameter's type, default and allowed values, so people running the report get a clear form and helpful errors.
gdotv also warns the agent if people with a read-only role would not be able to run the report, for example because it calls a procedure gdotv cannot verify as a read.
Add connections
An agent can add a connection for you. It tests the settings with gdotv's connection checks first and only saves the connection once the test passes, explaining what to fix if it fails.
Add a connection to my Neo4j database at neo4j.example.com, user neo4j. I'll give you the password.
On a gdotv server, connections are made from the server, not from your computer: localhost means the gdotv server itself and is refused, and command-line logins such as AWS CLI profiles or gcloud are not available. The agent checks which gdotv it is talking to before asking you for details.
See what agents have done
Activity, at the bottom of the side bar, lists what agents did while you were working: drafts created, edited, published or discarded, and connections added, changed or removed. The badge counts what you haven't looked at yet, and Open takes you to it. On gdotv Team and Enterprise, Activity also lists published dashboards and saved queries that other people changed or deleted.
gdotv never switches tabs on you when an agent creates something: it waits for you in Activity.
Links from your agent
Agents give you Open in gdotv links to their work: a result, the schema, a dashboard or a report. They can also hand you a query without running it, for example a change they aren't allowed to make, for you to review and run yourself.
- On gdotv Team and Enterprise, the link opens in your browser. It only works for your gdotv account.
- On gdotv Desktop, the link opens in the gdotv app.
- Links expire. If one does, ask the agent for a new one.
Tips
- Name the connection when you have several, to save the agent a step.
- State your assumptions: "active customers means an order in the last 90 days".
- Use the names in your data: label, relationship and property names lead to more accurate queries.
- Ask the agent to check itself: "count the rows another way to confirm the total".
- Prefer reports and dashboards for repeat questions: they keep working without the agent.