Chat

The Chat tab is the fastest path from “I saved this agent” to “I’m watching it answer.” It supports every feature of the runtime — multi-turn history, tool calls, structured output, attachments, voice — in a UI you can use without writing any client code.

​

Layout

• Sidebar — past conversations grouped under “Chats.” Click any one to resume; click New Chat to start fresh.
• Main chat area — message history with inline rendering of tool calls (each shows its name, arguments, and result).
• Input bar — the agent picker, version picker, context dialog, attachment buttons, and message input.

​

Picking an agent and version

• Click the agent name button above the input field to switch agents.
• Click the version button (defaults to Latest) to pin the chat to a specific version. Version 0 ($LATEST) is always the current mutable working copy; numbered versions are immutable snapshots. See Versioning.
• Picking an alias (e.g. production) routes each turn to whichever version the alias currently points at.

​

Adding context

Click the @ button to open the Add Context dialog:

• Context — a JSON object passed alongside each message. Useful for templated system prompts ({{user_name}}, {{org_id}}, …) or anything else your agent reads from the request body.
• Headers — a JSON object of HTTP headers added to each request — auth tokens, user IDs, anything you’d want a real client to send.

Both values persist for the rest of the conversation.

​

Attachments

• Files — paperclip icon. Attach PDFs, images, and other files inline with a message. The agent’s model needs to support the file type natively (e.g. vision for images, document inputs for PDFs).
• Audio — microphone icon. Record voice input directly. Useful for testing agents wired to the Transcript Generation tool.

​

Tool-call visualization

Every tool call the agent makes shows up inline in the chat as a labeled block:

• Tool name and source (e.g. Google Calendar: List Events, MCP: search_repo).
• Arguments the agent sent.
• The tool’s result.

Click any block to jump to its span in the full trace.

​

Sending messages

Type into Ask anything and press Enter (or click send). The response streams in token-by-token. Tool calls execute inline and stream their results as they complete.

​

Useful per-message actions

Each assistant message has a footer row:

• Copy the response.
• Thumbs up / thumbs down for feedback (logged for evals).
• Regenerate to re-run the last turn.
• View trace to jump straight to the OpenTelemetry trace for the run.

​

When to use Chat vs the HTTP API