All work with scripts and data happens through chat. You write to the agent in plain language — it replies, writes code, runs it and reports the result. The application works as a go-between: it shows the chat and the live stream of events, while the conversation itself is held by your central Xedant Agent.
Connection
The chat connects with two environment variables: AGENT_API_URL — the address of your Xedant Agent — and AGENT_API_KEY — its API key. Until they are set, the Chats section shows a notice that the agent is not configured; everything else in Data Agent keeps working.
How it works
The browser talks only to your own Data Agent server, and the server forwards requests to the agent, attaching the key on its side. The agent key never reaches the browser, and all chat features live on the external Xedant Agent — the application stores nothing of it and only displays it.
Live streaming
All chat events — new messages, processing stages, the queue’s state — arrive as one live stream of updates, without reloading the page. The agent’s reply “types itself” on your screen, and you see what it is doing: reading files, fixing a script, running it.
Models and skills
For every message you can pick a model (a fast one for simple questions, a powerful one for complex code) and a skill — the agent’s specialization. The list of models and skills comes from the agent itself and updates live: enable a new model in the agent’s panel, and it appears in the list at once.
Voice input
The input box has a microphone: press it, say your task, stop the recording — the text appears in the box, ready to send. This helps when it is easier to say than to type: listing fields, links, collection conditions.
Speech recognition is performed by your central Xedant Agent — the engine and language are configured there too. If the model is still loading, a hint about the preparation appears next to the microphone; after that recognition is instant.
The prompts queue
If the agent is busy, sent messages join a queue and show up as cards above the input box. Each card can be paused, edited or removed while the queue has not reached it. Tasks are worked through one by one, and nothing gets lost.
Git: every change under control
The agent commits every script change to Git. The chat panel shows the commit history, and uncommitted edits can be undone if something went wrong. This is the guarantee that any state of the scripts folder can be restored.
Licensing
Sending messages to the agent and managing the prompts queue require a valid license. Without one the application warns you: activate a license to send messages. Runs, the scheduler, the Lakehouse and browsing data work without a license. Details on the Data Agent Licensing page.
ETL sessions
When the agent performs a data collection task at your request, it opens an ETL session — a unit of work tied to the chat. Script runs made within the session carry the session’s badge and a link back to the chat: you can always see why a script was run and who ran it. More about the ETL cycle is on the Xedant Agent Integration page.