Agent actions
What the VizChat agent can actually do — edit your document, find and build widgets, read workspace files, search the web, and run code.
The VizChat agent is not just a "talking assistant" — it has a toolbox that changes your document directly. Knowing what it can reach helps you phrase a request it can carry out in one pass.
What the agent can do
| Capability | What it means |
|---|---|
| Edit the document | Read the document, make precise field-level edits, and validate the result |
| Find a widget | Search built-in and user-published widgets, and look up the fields and operations a widget accepts |
| Build a widget | When nothing existing fits, scaffold, write, compile, check, and publish a new custom widget |
| Read workspace files | List, read, and search documents you uploaded or created; create, rename, move, and delete them |
| Read the web | Search the web, and fetch the readable content of a page or link |
| Run code | Execute Python for calculation and data reshaping |
| Drive the UI | Open a document tab or show a notification in the app |
What a request turns into
"Make me a line chart of the last 12 months' revenue"
→ find a line-chart widget, then write it into the document
"Change the title of this pie chart to '2024 Channel Mix'"
→ a precise path edit on that widget's field
"Look at my uploaded Q3 report"
→ read the workspace document and pull out what it needs
"Search for recent AI funding news"
→ web search, then fetch the pages worth reading
"Compute the monthly growth rate for each region in Python"
→ run the calculation, then feed the result into a widget
"Build me a countdown timer"
→ build a new custom widget and publish it to your libraryOrder of operations
The agent generally works in this order:
- Read context — the current document, the widget you have selected, documents you referenced
- Gather material (when needed) — workspace files, web pages, computed data
- Find or build the widget — reuse from the library first, build a new one only if nothing fits
- Write and validate — write into the document, then validate the whole document in one pass
- Explain — say what was done, what the key trade-offs were, and what you might want to adjust
Confirmation
Two different things go by this name.
In conversation. Before doing something with a lasting effect — building a brand-new custom widget, for instance — the agent asks you in plain text and waits for your reply. Say yes and it proceeds; say no and it keeps discussing with you instead.
As a card. Deleting a file is the one action that cannot be undone from the conversation, so it does not rely on the agent asking nicely. When the agent tries to delete a document, the deletion is held and a confirmation card appears in the chat showing exactly which file is about to go. Nothing happens until you press Confirm; press Reject and the agent is told you declined and moves on.
Next step
BYOK (Bring Your Own Key)
Connect your own OpenAI / Anthropic / Google / DeepSeek / Qwen / Moonshot / xAI / Mistral API key — BYOK calls do not consume VizChat platform credits.
Workspace Overview
A workspace is the top-level container in VizChat that organizes canvases, storyboards, documents, and assets — one account can own many workspaces.