Overview

ObeliskDB is the agent-native data plane. A language model gives you an answer; ObeliskDB makes that answer defensible — governed, reproducible, permissioned, budgeted, and safe to write back — with the guarantees enforced below the model, where an agent cannot opt out of them.

The idea in one paragraph

When people wrote the queries, a data platform's job was to help humans express intent. Agents write the queries now, so the job inverts: make an autonomous, tireless, occasionally wrong author trustworthy. That trust cannot live in a prompt, because a prompt is advisory — it can be ignored, injected around, or lost between sessions. It has to be structural, enforced by the engine every statement passes through. That engine is ObeliskDB.

The full argument is in the whitepaper ("The Enforcement Boundary"). These docs are how you use it.

What you get

• - A real warehouse. A DuckDB execution kernel, immutable version-chained micro-partition storage, a catalog with time travel and zero-copy cloning, virtual warehouses with credits, and Snowflake-dialect SQL. See Architecture.
• - Obi, a conversational agent that queries data, builds things, and administers the platform — and everything it does is governed, budgeted, and staged for your review. See Obi & MCP.
• - The contract. Six guarantees plus authority-path controls the engine enforces on every caller, human or agent. See The guarantees.
• - Agent workflows. Writes stage on a branch for review, answers are pinned as insights with computed freshness, and metrics give one governed definition per quantity. See Agent sessions, Insights & memory, Metrics.
• - The rest of the stack. Ingestion, transforms, streams & tasks, apps, a Snowpark-compatible Python API, and a governance layer.

Quickstart

uv run obelisk ui         # the control plane at http://127.0.0.1:8501
uv run obelisk --help     # repl, mcp, build, ingest, cost, emulator, pgwire

Open the control plane and talk to Obi, or drop to the classic console for a SQL worksheet. Point an external agent at the platform over MCP, or connect any Postgres client over the wire protocol.

Obi uses the Anthropic API when ANTHROPIC_API_KEY is set, otherwise a local Ollama model — no API key required for the local path.

Two surfaces

• - Control plane — chat-first. Ask, review, approve; results render inline; agent changes appear as approval cards. Designed for the way people work with data when agents do the typing.
• - Classic console — worksheets, a SQL editor, the object explorer, and every operational tab. SQL is never required in the control plane, and never taken away here.

Where next

New here? Read Architecture, then The guarantees — together they are the whole mental model. Building an agent against it? Go straight to Obi & MCP and Agent sessions.