Core concepts

A handful of ideas run through the whole platform. Once these click, everything else in Donkit is a variation on them.

Agent

An agent is the unit you build, test, and publish. It's not a single prompt — it's a complete configuration that says how the assistant behaves: its instructions, the knowledge it can draw on, the tools it can use, the model that powers it, and the interface end users see. You shape an agent in the Builder and, when it's good enough, publish it as a product.

Your main assistant

You don't have to build anything to start. Every account comes with a main assistant — a general-purpose agent that researches, browses, writes and runs code, handles files, and generates media. What makes it more than a chatbot is that it knows your agents: when one of your agents was built for exactly that job or knows something it doesn't, it delegates instead of guessing and brings the result back into its own reply, and you can pin an agent to talk to it directly inside the same conversation. See Chat & your main assistant.

Configuration, not code

Under the hood, an agent is a declarative configuration. You never write the plumbing — retrieval, tool wiring, conversation handling — Donkit assembles it from your description. That's what lets a non-engineer build something that would otherwise take an ML team, and it's what makes an agent measurable and comparable: two configurations can be scored against the same dataset and the better one wins.

Versions & candidates

Every meaningful change produces a new version. The current version is what your users get; a candidate is a proposed replacement — from you, or from the optimization loop — that is adopted only if it measurably beats the current one. This gives you history, rollback, and a governed way to improve without breaking what works. See Versions & candidates.

Knowledge

An agent grounded in your material is far more useful than a generic chatbot. You upload documents into collections, connect them to an agent, and it answers from your content — with the retrieval configured for you. See Documents & RAG.

Capabilities (tools & skills)

Beyond answering from knowledge, an agent can act: search the web, run code, drive a browser, send email, generate media, query structured data, or call any external system through MCP. Reusable bundles of behavior are packaged as skills.

The canvas

Work that doesn't fit in a chat bubble — a report, a table, a page, a chart — lands on the canvas, the panel beside the conversation. Files an agent creates are collected there automatically, and the agent opens the one you should look at. Some of them are interactive: a page can put a button in front of you whose click becomes your next message. See The canvas.

Quality as a number

The idea that ties Donkit together: agent quality is measured, not felt. You define what "good" means with datasets, score the agent on dimensions, test it adversarially, and watch it in production. Measurement is what makes the optimization loop possible.

Two kinds of people

Donkit serves two audiences with one platform:

• Builders shape agents in the studio — describing behavior, connecting knowledge, running evaluations.
• End users simply pick a ready agent and ask. They never see prompts or pipelines. See Using an agent.

Administrators govern the workspace on top of both — who can build, who can use, and what the guardrails are.