// the find
rilldata/rill
The fastest business intelligence tool for humans and agents.
Rill is a code-first BI tool where you define models, metrics, and dashboards in YAML and SQL, then query them through a UI backed by DuckDB or ClickHouse. The local developer experience is genuinely fast — spin up, point at a Parquet file, get a working dashboard in minutes. The cloud offering handles deployment, sharing, and the conversational AI layer on top of your semantic layer.
The YAML-as-semantic-layer approach is the real win here: dimensions, measures, and time grains defined once, version-controlled, diffable in PRs, and now legitimately useful for LLM agents that can read and write the same files. DuckDB for local iteration and ClickHouse for production scale is a sensible two-tier engine story rather than forcing one size to fit all. The MCP server connecting agents directly to the semantic layer rather than raw SQL is a meaningful design choice — an agent querying `total_revenue` through a defined measure is more reliable than an agent writing ad-hoc aggregation SQL. CI pipeline is thorough: separate Go test, lint, proto-check, E2E web test, and migration validation workflows.
The 'agent-first' repositioning is mostly marketing — the underlying YAML/SQL model was already agent-friendly before they added `.claude/CLAUDE.md` scaffolding and called it a feature. The cloud paywall hits fast: local dashboards work but sharing anything requires Rill Cloud, and the pricing for teams is not cheap. 95 database migration files and provisioner code for Kubernetes, ClickHouse, billing, and River job queues means the operational surface is enormous for what should be a simple BI tool — self-hosting anything beyond local dev is a real commitment. The conversational BI / natural language query feature is real but the failure modes of NL-to-SQL on business metrics are well-known, and there's no indication of how it handles ambiguous or wrong answers gracefully.