Skip to main content
The Kodexa Agent is available in every project workspace. You do not install anything — once a companion runtime is configured for your organization or project, the chat pane is one click away.
If the floating Ask Agent button is not visible, no companion runtime has been configured for this project yet. See Configure the Companion before continuing.

​
Open The Chat Pane

There are two entry points inside the project workspace:
  • The floating Ask Agent button. A small circular button anchored to the bottom-right of the workspace. Click it to slide the chat pane in from the right.
  • The workspace sidecar. If your workspace layout already shows the chat pane docked, the agent is ready — just click into the message input.
The chat pane is dockable and resizable. You can keep it open while you work on a document, a task, or a resource editor.

​
Anatomy Of The Chat Pane

The pane has four parts you will use constantly:
  • Channel selector. At the top of the pane. Switches between conversations — typically the project channel, a per-task channel, and (for project admins) the admin channel.
  • Message stream. The conversation history for the selected channel. Agent replies are streamed token-by-token; tool actions and draft notifications appear inline.
  • Input box. Type your message at the bottom. Press Enter to send. Shift-Enter adds a newline.
  • Inline prompts. When the agent needs a decision, it can render option buttons or document pickers directly in the stream. See UI Interactions.

​
Your First Conversation

The agent already knows quite a bit about your workspace when you start typing. Try one of these:
The first two are read-only — the agent inspects your project and responds. The third is an authoring request — the agent will draft a change to the Invoice taxonomy and open the taxonomy editor so you can review the proposed edit before saving.
Speak naturally. You do not need to know tool names, resource UUIDs, or scheme prefixes like taxonomy://. The agent resolves names from context.

​
Starter Tiles And Pre-Built Prompts

You do not have to start from a blank input. When no conversation is selected, the chat pane shows starter tiles that match where you are: in a project, the tiles suggest project-level questions (summarize the project, explain the taxonomy); with a task open, they suggest task-shaped ones (summarize the task, explain the exceptions). Click a tile to start a chat with that question. Teams can go further and publish their own reusable prompts. When your organization has prompts for the current context, a From prompt button appears at the top of the chat list — it opens a picker of the organization’s Prompt resources, grouped by category. Picking one starts a new chat named after the prompt, with the prompt’s body prefilled into the input. Nothing is sent until you press send, so you can tailor the text first. Prompts are ordinary organization-level resources, managed through the standard /api/prompts API. Each one is shaped by a handful of fields:
  • name — the title shown in the picker, and the name of the chat it starts.
  • metadata.context — "task" or "project": which screen offers the prompt.
  • metadata.category — how the picker groups the prompt. Prompts without a category land under General.
  • metadata.prompt — the message body prefilled into the input. A prompt with an empty body is not offered.

​
What The Agent Already Knows

When you send a message, the agent automatically receives:
  • The project you are in (slug, name, bound resources).
  • The channel you are messaging on (task, project, project admin, or workspace global).
  • The task that is currently open, if any — title, status, linked documents, template fields.
  • The document you are looking at, if any — and which attribute is focused.
  • A snapshot of recent saves and active drafts in your session.
You do not need to repeat this context. Asking “What’s this task about?” in a task channel just works.

​
Asking The Agent To Make Changes

Most authoring requests follow the same loop:
  1. You describe the change you want.
  2. The agent inspects the relevant resource, drafts the change, and opens the matching editor tab.
  3. You see the proposed change in the editor — exactly as if you had typed it yourself.
  4. You review, tweak if needed, and click Save.
The change is not committed until you save. Closing the editor tab does not discard the draft — it stays in your session until you save it or explicitly discard it. For a deeper walk through this loop and the resources you can author, continue to Manage Project Resources.

​
Prerequisites

You need a few things in place before the agent can do useful work:
  • A companion runtime must be configured at the organization or project level. See Configure the Companion.
  • The user account you are signed in with needs access to the project — the agent acts as you, and inherits your permissions.
  • For authoring tasks, the relevant skills must be enabled on the runtime. If you ask the agent to edit a taxonomy and it tells you the skill is not enabled, ask your org admin.