The AI assistant
Verql includes a built-in AI assistant that can help you write SQL, understand your schema, and explain query results — using a provider and model you choose, with your own API key. It can also act as a tool for external AI clients through a built-in MCP server.
Providers and your API key
Section titled “Providers and your API key”The assistant supports these providers out of the box:
| Provider | Notes |
|---|---|
| OpenAI | Bring your own OpenAI API key. |
| Anthropic | Bring your own Anthropic (Claude) API key. |
| Ollama | Connect to a local Ollama instance — useful for running models on your own machine. |
You supply your own API key. Like your database credentials, keys are stored
encrypted in your operating system’s keychain via Electron’s safeStorage —
never in plain text on disk. (See
Connecting → How your credentials are stored.)
Chatting and tool calls
Section titled “Chatting and tool calls”You talk to the assistant in a chat panel. To actually help, it can use tools — for example, listing tables, describing a table, or running a query against your connected database.
Tool calls require your approval. The assistant doesn’t silently touch your database: each tool call is gated by per-call permission, and you approve it before it runs. This is especially important for anything that could modify data — write operations are not performed without your go-ahead.
Conversation history
Section titled “Conversation history”Your conversations are kept so you can return to earlier chats rather than starting over each time. You can revisit and continue prior threads.
The built-in MCP server
Section titled “The built-in MCP server”Verql also includes an MCP (Model Context Protocol) server. When you turn it on, it lets an external MCP client — such as Claude Code — connect to Verql and use a curated set of tools to read your schema and run approved queries against your active connection.
Key safeguards, all under your control in the MCP settings:
- You must explicitly start the server. It is off until you enable it.
- Tokenised endpoint. Every request must present a bearer token, so a random process can’t connect. (The server does not allow arbitrary web pages to reach it.)
- Per-tool permission gating. You choose which tools the server exposes with per-tool enable toggles — the same approval model the in-app assistant uses.
- Read-only mode. Flip this on to let external clients read without being able to write.
- Row limit. Cap how many rows a tool call can return.
- Live activity log. Watch what the server is doing in real time.
Because the assistant and the MCP server share the same underlying tool registry, a tool behaves consistently whether you invoke it in chat or an external client invokes it — including the write-approval gate. (For example, attempts to smuggle a data-modifying statement past the “read-only” label are routed through the same write-approval check.)
Only start the MCP server and connect clients you trust. Treat the bearer token like a password.
Next: Managing plugins →