[ docs ]product guide
vibecode your AI agent¶
Donkit is a platform where anyone can build an AI agent for their task — and their colleagues can use the result without knowing anything about RAG, prompts, or pipelines.
You describe how the agent should behave, connect your company's knowledge and tools, test it in a live chat, measure its quality on real examples, and publish it as a product with its own link. Donkit turns AI from a pile of one-off experiments into a governed, measurable capability that scales from a single builder to every employee.
Why Donkit¶
Most teams adopting AI get stuck between two dead ends. "Chat with your documents" tools solve exactly one scenario and give you no real control over quality. "Build it yourself" frameworks demand a team of ML engineers, ship no product interface for end users, and leave quality as a gut feeling rather than a number. Either way the project ends as a demo, not a product.
Donkit sits deliberately between "just another chat" and "assemble everything from open-source parts." It gives you a closed loop: build an agent → measure its quality → publish it for employees → gather feedback → improve.
-
Start by asking
Every account has a main assistant that researches, browses, writes code, and hands work to your own agents. Pin one and talk to it directly in the same chat.
-
Build by describing
Tell the Builder what you want in plain language. It writes the instructions, switches on the right tools, and generates a live interface you can try as you go.
-
Ground it in your knowledge
Upload your documents into collections and connect them. The agent answers from your material — you never configure retrieval by hand.
-
Give it real abilities
Web search, code execution, a browser, email, media generation, structured data, external tools via MCP, and reusable skills.
-
Measure and improve
Turn quality into a number. Build test cases, score the agent, and let the optimization loop keep only the changes that are actually better.
-
Ship it as a product
Publish an agent to a branded, shareable page — or call it from your own code through an API.
-
For builders and users alike
Builders shape agents in a studio; everyone else just picks a ready agent and asks. Admins govern who can do what.
What makes it different¶
Most tools let you tweak a prompt and hope. Donkit treats an agent as a declarative configuration whose quality is measured on datasets and whose changes are governed by versions and a promotion rule — a new version is accepted only if it measurably beats the current one. That turns "we fiddled with the prompt" into a repeatable engineering process with history and rollback.
In short: Donkit isn't just a runtime for agents — it's a platform for managing their quality over time.