Under the hood
The OpenClaw lineage
Built from OpenClaw's ideas, not its runtime — and where we diverged.
Dustav's architecture comes from OpenClaw, the MIT-licensed open-source agent runtime. The core ideas in this section — context as code, structured state you can read, negative space — are OpenClaw's. We took the ideas and built our own runtime on them: Dustav doesn't run OpenClaw's code, it runs on OpenClaw's thinking.
What we lifted
The load-bearing ones, almost verbatim:
- Context as code. What the agent perceives each turn is assembled, deliberately, from structured state plus live data — not dumped in wholesale. (Context engineering.)
- Structured, readable state. The agent's knowledge — the facts, the tables, the record of what it did — lives as legible records a person can read, not an opaque blob. That's the premise of everything on the record.
- Negative space. Half the craft is telling the model what not to do. (Negative space.)
If those feel like the spine of the whole product, that's because they are — and they're OpenClaw's spine first.
Where we diverged — on purpose
A homage that only flatters isn't worth much. The interesting part is where we deliberately departed, because that's where Dustav's own shape comes from.
- Operator of the agent → operator of a business. OpenClaw assumes a technically fluent user who lives in the files and runs the agent like a power tool. Dustav assumes a shop owner who will never open a terminal — so the state is visible but not hand-edited, and the agent makes its own edits by being told. Same transparency, opposite hands on the keyboard. (Read vs write.)
- A runtime → a mirror. OpenClaw is general; you point it at anything. We pointed everything at one job — the business inbox — and bet the product on a specific shape: an AI-native mirror of Gmail, where the store stays Google's, the primitives stay Google's, and everything proprietary lives in what the model understood. The narrow scope isn't a limitation we accepted; it's the moat we chose.
- Autonomy fenced by architecture. A power tool trusts its operator with every capability. An agent inside a stranger's business email must not trust itself that far — so the send-shaped powers were never granted, and the human commit is load-bearing structure.
Credit
The patterns are OpenClaw's; the decisions about who they're for are ours. If you build agents, OpenClaw is worth your time — it's where a lot of this thinking gets worked out in the open, and you can read every line of it.