Documentation
What Daslab is, how to set your first scene, and the concepts underneath — every page, grouped the way the sidebar reads.
Start here
- What is Daslab? — A workspace where AI agents do real work in the systems you already run. Writes wait for your approval.
- Get started — Run your first scene in five minutes.
- Your first blueprint — Write a scene as JSON and push it live.
- Daslab CLI and MCP — Drive Daslab from Claude, Claude Cowork, Claude Code, Cursor, the shell, or any MCP client.
- Vision — The same scenes, gates, and traces that run an inbox run a robot arm. The lab in Bangkok is where we test that on real hardware.
- React SDK — Every tool of every connected integration, as a typed React hook. Your first page needs no account and no API key.
- React in a cell — Build a cell with React from a CDN, no build step. The runtime hands you the data and the auth.
Concepts
- How Daslab works — One workflow walked end to end, from the scene it lives in to the run that gets recorded.
- Scenes — A shared workspace for one workflow, with live cells on top and the conversation that built them below.
- Assets — An asset is a thing you would miss. Your Gmail account, an invoice record, a note the agent keeps.
- Cells — A cell is how you look at an asset. It holds no truth of its own, so rearranging cells never needs approval.
- Knowledge Bases — Notes with folders and links, kept in the scene that uses them, where the agent can search and quote them.
- Agents — The part that reads the prompt, picks the tools, and decides what to do next. You choose the model.
- Jobs — One conversation that ran: the prompt, the tool calls it triggered, the result, and the trace.
- Layers — Every asset and every job sits in a layer: Private, Shared, or Public. Where a thing sits is who sees it.
- Identity — People and devices hold keys. Agents hold nothing: every run borrows its authority through a grant, and history records the chain.
- Scripts — The plan as a value you can read before it runs and re-read after. Steps can be addressed to people, and can wait on the world.
- Approvals — Reads run freely. Every write stops at a gate, with the exact diff, before it commits.
- Sandbox — When no tool fits, the agent writes code and runs it in an isolated sandbox that can't reach your systems.
- Providers — GitHub, Gmail, Postgres, S/4HANA. Each connects as a provider, bringing its asset types and the tools that act on them.
- Blueprints — A scene written as one JSON file. Apply it and it appears; export a live one and you can fork it.
- Provider Scenes — Build an integration by filling a scene with typed rows instead of writing a TypeScript module.
- Commit chain — Every scene change lands as a commit on a hash-linked chain, so “what changed since Tuesday” has an exact answer.
- Traces — Each job leaves an OpenTelemetry trace: what ran, in what order, what it cost, and what came back.
- Audit — What the traces and the commit chain log, who may read them, and how long they are kept.
Operating
- Auth & permissions — How you sign in, how Daslab connects to your tools, and why the agent can never do more than your own account can.
- BYOK (bring your own keys) — Managed usage by default: one bill, hard caps. On Enterprise, run inference on your own provider accounts.
- Data, residency, retention — Where your data lives, how long it stays, and why we don't train on it.
- Microsoft 365 permissions — Exactly what Daslab asks for in your Microsoft tenant, what it deliberately does not ask for, and how to grant or revoke it.
- Daslab Gateway — A read-only copy of a system that cannot leave your network, answering questions from your scenes over one outbound connection.
Reference
- API console — Interactive OpenAPI reference for the REST API.
- API v1 reference — The current API surface, including OpenAI-compatible jobs.
- Self-hosting — Run your own instance — one binary, a Postgres, a Redis.
Influences
- Influences — Alan Kay, Bret Victor, Tufte, Apple's WidgetKit. This page cites what we took from each, and what we decided not to follow.
Changelog
- Docs site, take one — We rebuilt the website around docs, and the old marketing pages are gone.