The AI assistant
A panel in the window that answers questions about your job.
Last updated November 19, 2026
- Pick the agent
- What it sees
- What it does
- Create a connection
- What it does not do
- A local model
- Questions to ask
The assistant is in DBConvert Studio only. The pair tools do not have it.
The assistant is optional. Until you pick an agent in Settings, then Assistant, it does nothing and sends nothing anywhere.
The assistant reads your job and what DBConvert answers about it. It explains. You decide and press the buttons.
It runs on a local model (Ollama) or on an agent you already have: Claude Code, GitHub Copilot CLI, Codex, Gemini CLI, OpenCode or Goose, on your own account with that agent's provider. You pay your provider. DBConvert charges nothing for it.
Pick the agent
Open Settings, then Assistant. One line per agent. The picked one opens and shows its description and its fields:
- Local model (Ollama): a model on your own Ollama. See A local model.
- Claude Code: Anthropic's agent, on your Claude account.
- GitHub Copilot CLI: on your Copilot plan.
- Codex: OpenAI's agent, on your ChatGPT or OpenAI account.
- OpenCode and Goose: open-source agents, on a model provider of your own.
- Gemini CLI: Google's agent, on your Google account.
- No agent is picked at first. The panel says to pick one.
- An agent that is not installed shows Not installed and a link to its install page.
- One agent is used at a time. A change applies from the next conversation (New chat).
- Codex, OpenCode, Goose and Gemini CLI carry the tag Not tested yet. Before the first use, a notice says: not tested with DBConvert yet; the agent may use its own tools on this computer, such as reading files or running commands; each such request asks you first. Click Use and the agent's name to go on, or Cancel.
- Claude Code and Copilot CLI have no tag. Claude Code runs with its own tools switched off. Copilot CLI is started to see DBConvert's tools only.
Choose the model
- Claude Code: Claude Code's own model (the default, as set in Claude Code), Haiku, Sonnet or Opus.
- Codex: before its first conversation the field says that Codex lists its models after its first conversation. After that, the list has Codex's own and the models Codex offers.
- Copilot CLI, OpenCode, Goose, Gemini CLI: no model field. The agent uses its own setting.
The model applies from the next conversation.
Claude Code needs Node.js and Claude Code installed, and signed in. Its first start downloads about 0.5 GB. Copilot CLI needs GitHub Copilot CLI installed and signed in.
What it sees
It never sees your passwords.
It sees rows of your tables only while Show sample rows to the assistant is ticked. The box is under the question field and in Table from SQL. It is one setting for the agent in use:
- for an agent other than the local model, it is off each time the window opens;
- for the local model, it is on until you untick it, and your choice is kept.
Otherwise it sees:
- names of connections, tables and columns;
- host, port and database name, when it checks a connection;
- column types;
- counts of rows and tables;
- the messages DBConvert and the database servers return.
Error messages reach it without values from your tables.
More on what leaves your computer: Security and privacy.
Passwords stay in the window. The assistant never sees a password and never asks for one. Do not type a password in the chat.
What it does
- Reads your job. Source, destination, the selected tables, options, the last run and the last Validate data result.
- Explains a table or a column. Its types, what is lost, the warnings and what Validate data does not compare.
- Explains a run's errors. It groups the errors across tables and puts one table's error next to its column types.
- Names the risks of a plan before Start conversion. Tables that will be replaced, emptied or added to, values that lose something, columns that Validate data does not compare, writes to the source. It also names the tables left out: Blocked (a blocking problem leaves them out of the run) and Unselected (unticked in the job, so the run does not copy them).
- Checks a connection. It shows the step that failed and what to fix. You fix it in the window, then it checks again.
- Builds the plan. The same plan as Plan conversion. Nothing is copied.
- Includes or leaves out the tables you name. It also sets what happens to tables that already exist in the destination: append, empty first or overwrite. The change is kept in the job at once, and the window shows it. You choose what happens to existing tables; it does not pick for you.
- Writes the SELECT of a Table from SQL from your words, and fixes it when the check fails. It also counts the rows of the result when you ask. You press Undo to put back what the fields held.
- Creates a connection from your words. See Create a connection.
- Opens a place in the window. A step, a table, the issues on Tables & mapping, Validate data, Connections or Jobs.
It answers only from what DBConvert returned. When there is no answer, it says so and does not guess.
Create a connection
Say which database to connect. It works with any agent.
- Write what you know, for example: "Connect the source: MySQL on 10.0.0.5, user root, database shop".
- The assistant opens the New connection window and fills in what you named: type, host, port (the default, if you did not name one), user, database and name.
- If you did not say source or destination, a card in the chat asks: Source or Destination.
- Type the password in the Password field of that window. The cursor is already there. The window says: Enter the password, then Save.
- Click Save.
For a cloud database (Neon, Supabase, Amazon RDS and others), the assistant opens the Connection string tab. Paste the string from your provider's console.
The chat says: Don't type passwords in the chat: enter them in the connection window.
What it does not do
- It does not approve a plan.
- It does not start, stop or schedule a run.
- It does not save a Table from SQL: Add to job is yours.
- It does not change a saved connection.
- It does not change column types, options or schedules.
- It does not delete anything.
Asked to start the transfer, it opens Execution. You press Start conversion.
Questions to ask
- "Why did the run finish with errors?"
- "What will this conversion do to the tables that are already in the destination?"
- "The source connection does not connect. What is wrong?"
- "Connect the destination: PostgreSQL on db.example.com, database shop."
A local model
With Local model (Ollama) the assistant runs on a model in your own Ollama. In Settings, Assistant:
- Set the address of Ollama. The default is
http://localhost:11434. - Pick a model from the ones Ollama has. When you pick Local model and no model was picked before, it is
qwen3:8b. A model you picked stays. - Click Check connection. It shows Checking… while Ollama is asked.
- The assistant sends its requests only to the address you set. With Ollama on your computer or in your network, nothing reaches the internet.
- Each request carries the names and types of the tables and columns of the source (up to 40 tables). For the rest, one line says how many more tables there are, and the model looks them up with its tools. While the sample-rows box is ticked it carries up to 3 rows of each table.
- While it works, the panel shows the model's thinking as it comes, if the model can think, and the answer as it streams in.
- The panel says why it cannot start: Ollama is not running at the address, no model is picked, or Ollama does not have the picked model (run
ollama pullwith its name, for exampleollama pull qwen3:8b).
The model's log
The assistant writes what the model was given and answered, step by step, to local-model.jsonl in %LOCALAPPDATA%\DBConvert\Studio\logs\assistant\. It holds table and column names, the answers of DBConvert's tools and, while the box is ticked, sample rows. It holds no connection password.
Attach it to a message to support only when we ask for it, and read it first: it can hold your data.