Talking to the agent: text, voice, images and pointing

The prompt bar sits at the bottom center of the workspace. It has a sparkle icon, a text field, an image-attach button, a microphone, a Smart / Fast switch, an effort control, and a round coral send button. Everything you want from the agent starts here.

There are four ways to talk to it. They are not exclusive. The best requests often mix two or three.

1. Type it

Plain language works. Press Enter to send and Shift + Enter for a new line. Prompts can be up to 4,000 characters, which is plenty.

Typing is best when you are giving precise numbers, naming parameters, or pasting a measurement list.

Make the wall thickness 2.4 mm and add a 3 mm radius to every outer vertical edge.

While the agent is working, the send button turns into a Stop button.

Leaving while it works: If you click another tab (Concept, Make, History), Home or any other link while the agent is still working on your model, Solid Vibes asks "The AI is still working on your model, leave anyway?". Stay is the default. Leave stops the request: the preview it was building is dropped, your model stays as it is and nothing is added, so you ask again when you are back. Closing or reloading the browser tab asks the same way. Once the agent has finished, you can move around freely.

2. Speak it (push-to-talk)

Click the microphone, or press and hold Space when you are not typing in the prompt bar. Speak your request. Live captions appear under the prompt bar while you talk, so you can see what was heard.

Push-to-talk is the fastest way to give one clear instruction, especially when your hands are on the mouse.

Dictation costs 1 credit per minute by default, plus whatever the resulting request costs. A spoken request goes through the same preview-and-accept path as a typed one.

3. Talk it through (Live voice mode)

Live voice mode is a real back-and-forth conversation. You speak, the agent answers out loud, you interrupt, ask "why did you do that?", and steer as you go. It works with the same agent behind the scenes. Anything it builds still arrives as a ghost preview, and your current selection is shared with it, so "this" keeps working.

Live voice is metered by minutes per month, and the allowance depends on your plan (the free trial includes 15 minutes). It uses more credits per minute than dictation: by default 4 in Fast mode and 6 in Smart mode.

Use it when you are exploring or thinking out loud. Use push-to-talk or typing when you know exactly what you want.

4. Point and say "this"

Click a face, an edge, a whole piece, or a point on the model, then say or type "this" or "here".

Make this 3 mm thinner.

Put a 5 mm hole here, 10 mm from the edge.

The agent receives a description of what you selected (for example "top face, 80 by 50 mm") and resolves "this" to it. Select several things at once and say "these". There is more about this in Selections and "this".

Add images

Click the image button or drop files onto the prompt bar. You can attach up to six images per request. Solid Vibes classifies each one automatically and shows its confidence:

Anything written inside an image is treated as data, never as instructions. The agent will not obey a note that says "ignore your rules".

The virus check

Every picture is checked for viruses before the agent looks at it. While that happens the attachment says Checking the file… and the send button waits; it usually takes a second or two. Nothing that hasn't passed the check is ever sent to the AI.

• If the check is slow, the agent waits a short while for it.
• If it still hasn't finished, the attachment says We couldn't check your picture for viruses yet — try again in a minute. Nothing was sent and no credits were used. Remove it and attach the picture again in a minute.
• If the scanner flags the file, you are told it was flagged and removed. Export the picture again from the original program, or use a different one.

Ask it things, too

The prompt bar isn't only for changes. Ask "how tall is it?", "how much does it weigh in PETG?" or "is this printable without supports?" and you get a plain answer measured on your real model — nothing changes. When the answer points at a fix, it ends with a one-click button such as "Make the walls 2.5 mm thick?". See Ask questions about your model.

The agent talks back

The line above the prompt bar shows the latest reply. The whole conversation — every answer in full, every preview and what happened to it — is in the Chat panel on the right, which opens by itself the first time the agent replies. See The Chat panel.

Sometimes the agent asks a question instead of guessing: a choice between options, or a field for a missing number such as the diameter of a pen. This happens when its confidence is low. Answering is cheaper than undoing a wrong guess.

You may also see small dismissible nudge chips like "Fillet all 12 sharp outer edges?" These are suggestions only. Ignoring them costs nothing.

Sometimes the agent also offers a way to look at the result somewhere else, for example "Want the full view?" with an Open Check (FDM) button, or it opens a Make page you asked for ("export this as STL"). Those never appear while the agent is still working: they wait until it has finished, so you are not pulled away from the build. Press Not now to dismiss the offer.

Pick the right channel

Try it

On any model, select one face and say, or type:

What is the area of this, and how far is it from the opposite face?

You will get a measured answer on the model itself. That is the "Ask, don't measure" idea: you can interrogate your design without touching a measuring tool.