The AI assistant
The assistant is not a chat beside the platform but an agent inside it. It uses the same paths as the interface — with the same rights and the same logging.
On this page
Pyro in context
Pyro can show charts in chat, turn them into dashboard tiles and create or change lavaflow drafts. A Pyro transformation saves generated Python after your review for later runs. The separate Pyro step for classification, extraction and summarization instead calls a local or API-backed model on each run.
It acts with your rights
The agent holds no privileged account. What it queries, it queries with the token exchanged for you. A row a policy hides from you is hidden from it as well — and it therefore cannot mention that row in an answer.
Gate, audit, run
The agent does not call a database directly; it works through a fixed set of curated tools. Each of them passes the same three-step chain:
Gate
May this user call this tool with these arguments? The check is fail-closed: whatever is not explicitly allowed is refused.
Audit
The call is written into the hash chain before it runs — not after.
Run
Only now does the tool execute, under the delegated token.
Mutating steps ask first
Anything that creates, changes or deletes is surfaced for confirmation first. The agent acts once somebody agrees — not because it thinks it should.
What it is good for
- Looking at and analysing data — walking the catalog, building queries, explaining results.
- Proposing and creating the semantic model: it can inspect a table and derive metrics and views from it. Write-gated and administrators only.
- Operating dashboards — viewing, changing, renaming and deleting tiles.
Curated metrics are queried through stored definitions. The agent can also manage report and threshold-alert schedules and create and manage working copies. Available tools determine which setup steps are possible in chat; not every console endpoint is exposed as an agent tool yet.
The language model is swappable
No model ships with the platform; administrators choose the backend. Three routes are open: a key for a hosted provider, a self-hosted model behind an OpenAI-compatible address, or letting individual users sign in with their own Claude account. Details under Configuring the AI backend.
The AI backend determines the data path
External models receive conversation content and tool results used in chat. Optional classification sends only column names and types to the model. A self-hosted language model and local embeddings keep this processing on your own network.
lavalake as an MCP server
The reverse direction works too: the platform exposes a selection of read-only tools over the Model Context Protocol so that other AI systems can query the lakehouse. Enablement is fail-closed — a tool is reachable only once it has been explicitly switched on.
Applies to: This page describes lavalake 0.201.0. Product and documentation live in separate repositories — when in doubt, what the console shows is authoritative.
Something missing?
If these pages do not answer your question, we will walk through the platform against your own case in half an hour.
Book a demo