AI agents are starting to query, change, and spend against company data. Today the only thing keeping them in line is a prompt they can ignore. ObeliskDB is a data plane that enforces the rules itself — who sees what, what a metric means, how much an agent may spend, which writes ship — so every agent, from any vendor, is governed whether it cooperates or not.
DuckDB kernel · versioned storage · Snowflake-dialect SQL · MCP-native
The layer beneath your agents where the data actually lives: Snowflake-dialect SQL on a DuckDB kernel, with versioned storage. The rules for agents — permissions, budgets, metric definitions, write review — are enforced by the query engine, not by the agent's prompt. Obi, the built-in agent, and any outside agent over MCP get the same contract.
Databases were designed for people who wrote queries and could be trusted to read a policy. Agents are tireless, fast, sometimes wrong, and steerable by whatever text they read. A system prompt is advice. A permission dialog is advice. A rule the engine enforces is not.
Every answer comes with a receipt pinned to a data version. Every agent write lands on a branch until a person approves it. A spent budget refuses the next call. A masked column never reaches the model. Governance that works even when the agent does not cooperate.
When people wrote the queries, the platform helped them express intent. When agents write them, the job is to make an autonomous, tireless, occasionally wrong author trustworthy. That trust cannot live in a prompt.
The same rule is ignorable in the prompt and non-negotiable in the engine. ObeliskDB puts it in the engine.
Bring any model — Obi, an external coding agent over MCP, whatever comes next. Each inherits the same contract, because it is enforced beneath the model and cannot be opted out of. A masked column's bytes never reach the model. A spent budget refuses the next call. A write lands on a branch until a human merges.
One canonical definition per quantity. An agent selects it by name and cannot misdefine revenue.
Every agent write auto-stages on a zero-copy branch. A human reviews the row-level diff and merges. The pull request, for data.
Every answer replayable and pinned to the exact data version it was computed from. Reproducible forever.
Credits charged at execution and refused when spent. A runaway loop cannot spend your month.
Stored answers with freshness computed from the version chain — current, stale, or invalidated — and governed on read.
Policies applied inside the query rewrite, so filters see masked values and small-cohort probes are refused.
Ask questions, request changes, review what agents did, approve what ships. Obi runs on the Anthropic API or a local model — and everything it does is governed, budgeted, and staged for your review. SQL is there when you want it, never required.
A vendor-neutral position paper on why the guarantees an agent economy depends on must live in the data plane — with a reproducible evaluation and the market reason the incumbents are structurally disinclined to build it.
ObeliskDB is the fixed, authoritative thing beneath your agents — the source of truth they orient by and cannot move.