AI agents

Ask questions in plain language and get governed answers, charts, and dashboards built from the metrics in your Lightdash semantic layer.

AI agents are available as an add-on for all plans. View pricing

Lightdash AI agents let your team ask questions in natural language and get answers built from your semantic layer — the saved dashboards, metrics, dimensions, joins, and descriptions you’ve already defined. An agent picks the relevant models and metrics, builds and runs the query with the right filters, and returns the result as the chart, table, or summary that fits the question. Because every answer runs through the semantic layer, it respects the same project permissions and user attributes as the rest of Lightdash.

You can run agents in the Lightdash app or in Slack, scope each one to a domain with tags, and improve them over time with verified answers, evaluations, and knowledge documents. The pages below cover setup, day-to-day use, governance, and the features that make agents more accurate.

Looking for Autopilot? Autopilot is a separate, admin-only agent that runs on a schedule to keep your project clean — fixing broken charts, flagging stale content, and suggesting new ones. It does not build charts or dashboards from user questions like the conversational agents described here.

Choosing your AI surface

Lightdash exposes several AI surfaces, and the right one depends on who is asking, what they’re trying to do, and how much governance you need. They share one foundation: the Lightdash semantic layer — governed metrics, joins, descriptions, permissions, project context, verified answers, and evaluations. Investing there pays off no matter which surface a user picks.

Decision guide

Pick the first item that matches what you’re trying to do:

• Let business users ask governed questions in natural language — use AI agents in the app, or in Slack when the answer belongs in a team conversation.
• Query Lightdash from Claude, Cursor, Codex, ChatGPT, or a custom agent — use the Lightdash MCP server.
• Give your agents tools from Notion, Linear, Confluence, or other services — connect external MCP servers.
• Teach your AI coding tool Lightdash concepts and YAML syntax — add the Docs MCP and install agent skills.
• Build an interactive, shareable app from a prompt — use data apps.
• Create or edit saved charts and dashboards conversationally — use content editing. (Beta)
• Make bulk, reviewable dashboard changes — use dashboards as code with lightdash download and lightdash upload.
• Change a metric, dimension, join, or description — edit dbt or Lightdash YAML through Git, or use AI writeback. (Beta)
• Fix something the agent got wrong — improve descriptions, AI hints, project context, verified answers, evaluations, or Reviews findings. (Reviews is Beta.)

Compare the surfaces

Surface	Best for	Typical user	Governance model	Learn more
AI agents (app & Slack)	Governed self-serve questions and visualizations	Business user	Agent tag scope, project permissions, user attributes	Using AI agents, Slack
Lightdash MCP server	Querying Lightdash from external AI clients	Power user, developer, analyst	OAuth, project permissions, user attributes	Lightdash MCP server
External MCP servers	Giving your agents tools from outside services	Agent builder	Per-project config, per-user or shared OAuth	External MCP servers
Docs MCP + agent skills	Giving coding agents Lightdash concept and YAML context	Developer with an AI IDE	Public docs, local to the developer	Docs MCP, Agent skills
Data apps	Interactive, shareable apps generated from a prompt	Analyst, app builder	Project role and space access; queries run through Lightdash	Data apps
Content editing (Beta)	Conversational chart and dashboard creation and edits	Analyst, agent operator	Enabled per agent, scoped by agent permissions	Content tools
Dashboards as code	Bulk, reviewable, version-controlled dashboard changes	Developer, analytics engineer	Git review, previews, validation	Dashboards as code
AI writeback (Beta)	Agent-proposed changes to dbt / Lightdash YAML	Analytics engineer	Feature flag, Git review, dbt / Lightdash validation	AI writeback

A few principles hold across all of them: keep agents focused — several domain-scoped agents beat one project-wide agent; treat raw SQL as an escape hatch and fix the semantic layer when routine questions need it; and never upload sensitive material as knowledge documents unless everyone who can use the agent is allowed to see it. For the full treatment, see Effective analytics with agents.