The Kit Assistant

The Kit Assistant is the chat in the AI Dev Kit dashboard, at http://localhost:4600/#chat. It answers questions about your conversions and about Unyson+ itself, using the kit's own tools: it can open a web page and read it, pull up a capture's conversion report, search the kit's documentation, and — with your confirmation — start a capture.

It runs on whichever AI backend the kit is set to, so the conversation stays on your machine (a local model through Ollama) or goes to your own Claude subscription or API key. Nothing is sent anywhere you have not configured yourself.

:::note Not a general coding agent The assistant has seven tools and no ability to edit files, run commands or change your site. It reads things and answers. Anything that changes something waits for your Confirm — see What it can do below. :::

What it can do

Tool	What it does
look_at_url	Opens a page in the kit's own browser (real Chrome, so JavaScript-built sites work) and reads its title, headings, navigation and visible text. Leaves a screenshot.
conversion_report	A capture's report: per-section decisions, which elements fell back and why, dropped styles, style coverage.
search_docs	Searches the kit's documentation — section recipes, shortcode and option references, Theme Settings, the build protocol.
read_doc	Reads one kit doc in full.
capture_status	How a capture is going, or how it ended.
start_capture	Asks first. Proposes a capture; nothing runs until you press Confirm.
hand_to_studio	Asks first. Proposes handing a scene request to Kinetic Scenes.

Anything that writes comes back as a proposal — what would happen, where, and what it would replace — with Confirm and Cancel buttons underneath. The assistant cannot press them, and a proposal expires after 30 minutes. This is why it will say "this needs your Confirm" rather than claiming it has started something.

The conversation is kept in capture-out/chat-history.json (the last 200 messages) so it survives restarting the service. New chat clears it.

Attaching an image

You can show the assistant a picture — a screenshot of a page that came out wrong, a design mock-up, a photo of a whiteboard. Three ways:

• the paperclip button in the composer,
• paste an image straight into the message box,
• drag and drop one anywhere onto the composer.

Up to four images per message, in PNG, JPEG, GIF or WebP. Each one appears as a thumbnail above the box before you send, with an × to remove it, and stays visible in the conversation afterwards.

Large images are scaled down in your browser before they are sent (to 1568px on the long edge, which is as much as the models actually read). Screenshots stay PNG so their text stays crisp; photographs are re-encoded as JPEG. A GIF is sent untouched so its animation survives. The images themselves are saved under capture-out/_chat/ and the conversation stores only a reference to the file, so your chat history does not balloon.

Your model has to be able to see

Reading an image is a capability of the model, not of the kit:

• Claude (either backend — your subscription through Claude Code, or an API key) reads images.
• A local model reads images only if it is a vision model. In Settings → Local AI models these are marked with a blue vision badge; Qwen3-VL 8B is the recommended one, and Gemma 3, Mistral Small 3.2 and the other Qwen3-VL sizes also work.

If you attach an image while a text-only local model is picked, the assistant says so and names a model that would work, rather than answering as though it had seen the picture.

Commands

Typing / in the message box opens a list of commands. Each one runs a single tool directly, with no model involved — which makes it instant, exact, and equally reliable on a small local model that might otherwise need persuading to pick the right tool.

Command	What it does
/look <url>	Open a page in the kit's browser and read it
/report [capture]	The conversion report — what fell back, and why
/status [capture]	How a capture is going, or how it ended
/docs <query>	Search the kit docs
/doc <path>	Read one kit doc in full
/capture <url>	Capture a site — asks you to Confirm first
/studio <request>	Hand a scene to Kinetic Scenes — asks you to Confirm
/clear	Clear the conversation
/help	List the commands

Where a command takes [capture] you can leave it out and the latest capture is used, or name a capture folder, a URL, or latest.

Use ↑ / ↓ to move through the list, Tab or Enter to pick one, Esc to dismiss it. A command that needs no argument runs as soon as you pick it. Commands that change something still go through the same Confirm step as the assistant's own proposals — /capture proposes a capture, it does not start one.

Anything that does not begin with / is an ordinary message to the model. If you attach an image, it is always sent to the model, since looking at a picture is not something a single tool can do.

Limits

• Seven tools, nothing else. It cannot edit a file, run a command, change your WordPress site or touch git.
• The last 12 turns of the conversation are sent to the model; older ones scroll out of its view (they stay in the transcript).
• Six tool rounds per message. After that it answers with what it has gathered.
• It is for this dashboard. The chat endpoints refuse requests from any other web page.

If it says AI is off

The chat needs a model. Open Settings, and either pick a local model under Local AI models (the kit downloads it for you) or enable Claude — the kit finds the claude CLI on your PATH, or an ANTHROPIC_API_KEY if you have one set. The chat header always names the model that is answering.