The admin chat

The Chat pane lets you talk to Darwin about your site. Use it to look things up and make small edits, the kind you would make in wp-admin yourself. It never changes anything until you click Apply. Bigger or technical work goes to your own agent through a handoff.

Setting up the chat

Until a model is connected, the Chat pane says To chat here I need an AI model. Pick how. and offers these choices:

• Connect a provider (or Finish connecting): use Claude, Gemini or OpenAI with an API key. An API key is the password-like code a provider gives you for paid access to its AI.
• Work with your own agent: skip the chat and create a handoff for Claude Code or a similar agent.
• No API key? Use a local or self-hosted model: a model on your own machine or server, or an AI your host runs for you. These connect through a connector plugin for OpenAI-compatible servers.

Connecting a provider

1. Click Connect a provider. This opens Settings > Connectors, where the Darwin AI panel shows three steps that tick themselves off.
2. Provider installed and activated: install and activate the provider's connector from the list on that screen.
3. API key added: paste your key into that provider's card on the same screen. A provider that needs no key shows Connected instead.
4. Provider answered a ping: click Ping next to the provider in the panel. Darwin sends it one short test message. When it answers, the chat is ready, and you are taken to the dashboard with the panel open.

If you connect more than one provider, the chat uses the last one that answered a ping. You can ping any connected provider again from the Connectors screen at any time.

Models

An administrator picks the model under Settings > Global > Chatbot > Model. The choice applies to the whole site, for every chat and ping with that provider. Each model's list price per million tokens is shown next to it. A token is a small chunk of text, roughly three-quarters of a word.

Provider	Models offered	Default
Anthropic	Claude Haiku 4.5, Claude Sonnet 5.5, Claude Opus 5.5	Claude Haiku 4.5 (marked Recommended)
Any other connected provider	The provider's own default model	Not selectable in Darwin AI

For a provider without a model list in Darwin AI, usage is counted in tokens only, with no cost estimate.

Spending limits

Each chat message can cost money with your AI provider. Darwin AI has two budgets that pause the chat once they are reached:

Budget	Default	What happens when it is spent
Hourly budget	$1.00, on	The chat pauses until the next hour.
Daily budget	$5.00, on	The chat pauses until tomorrow.

• Change them under Settings > Global > Chatbot. Only administrators who can manage site settings can change them. Each budget can be switched off or set anywhere from $0.01 to $10,000.
• Hours and days follow your site's timezone. Budgets apply to the whole site, not to each user. Pings count too.
• The Spending meter shows this hour, today, and today per model, with request and token counts.
• The figures are an estimate. They are calculated from the tokens each answer used, at the list prices in the plugin. Your provider's invoice is the true cost. Two requests finishing at the same moment can go over a budget by one request.
• Budgets count dollars, so they only limit models that have a listed price.

What the chat can do

The chat uses Darwin AI's tools, but only a fixed, smaller set of them. Tools for plugins you do not have, such as WooCommerce, only appear when that plugin is active. In plain terms, the chat can:

• Pages and posts: find, read and search content; create a post or page, duplicate one, find and replace text; change status; move to the bin, restore, or revert to an earlier version.
• Elementor: read a page's outline, elements, settings, templates and form submissions; change the text of existing elements.
• Media: list and inspect files and where they are used; upload a file; edit details such as alt text.
• Menus: read menus; create a menu, add items, assign it to a location.
• Widgets: read widget areas and widgets.
• Site: read site information, settings and users; change general settings (such as title, tagline, timezone, reading and discussion defaults, permalinks) and the front page.
• Plugins and theme: list plugins and see their details; activate or deactivate a plugin; read basic theme and must-use plugin information.
• ACF (Advanced Custom Fields): read field groups, options pages, post types, taxonomies and values; create post types and taxonomies; set field values.
• Rank Math: read SEO audits, page SEO, schema, settings, redirections, 404 log, sitemap and IndexNow status; edit SEO for a post or term; add a redirection rule.
• WooCommerce: read store status and settings, products and orders; create or edit a product, and update stock.
• LiteSpeed Cache: read cache status; clear the whole cache, one post or one URL.
• Defender: read security status, audit log, lists, lockouts, scans, tweaks and two-factor settings.
• WP Mail Logging: read sent emails and email statistics.

Your Chatbot permission level narrows this further. It is set under Settings > Personal > Chatbot, and the default is Editor. Levels work the same way as for handoffs. At Editor, for example, the chat cannot change site settings, read the user list or orders (Manager), or activate and deactivate plugins (Owner). Raise the level only when you need it.

What it refuses, and the handoff card

Some work is deliberately left to your own agent, whatever the chat's level. That includes:

• designing, building, restructuring or restyling pages;
• theme, plugin or must-use plugin code, and CSS;
• files, migrations and plugin updates;
• deleting things permanently, and plugin settings.

Every change also has these limits:

• A single change may carry at most 4,000 characters of text.
• It may not contain layout markup, such as sections, styles, classes or layout blocks.
• On Elementor pages, it may only change the text of existing elements. It cannot move, remove, restyle or add them.

When a request needs your agent, the chat shows a handoff card with a short brief and the permission level the task needs:

• This needs your agent: the chat will not attempt the task.
• Your agent is better suited for this: the chat could do it, but the job is long (lots of writing, many pages, a whole-site audit). The card adds a Do it here anyway button.

Create handoff opens your own new-handoff form, with the name taken from the brief and the level preselected. Copy copies the brief to paste to your agent. The chat can never create a handoff itself.

Apply or Decline

Reading happens straight away. A change never happens on the chat's say-so:

1. When the chat wants to change something, it stops and shows a card marked Waiting for you. Where the tool supports a practice run (called a dry run), the card shows exactly what would change. It also lists the main details of the change.
2. Click Apply to make the change, or Decline to drop it.
3. The card updates to Applied, Not applied or Declined, and the chat tells you what happened.

If you send a new message while a change is waiting, that change is treated as declined. Every tool the chat runs is recorded in History under the name "chatbot".

Built-in limits

• Up to 6 rounds of tool use per message. If it needs more, the chat stops and asks you to try again more narrowly.
• Each tool answer passed to the model is cut to 6,000 characters.
• Each reply is capped at 2,048 output tokens, and each request may take up to 120 seconds.
• Your conversation is kept at about 16,000 tokens. Older exchanges are dropped first, which keeps each request's cost bounded.
• Each user has one conversation, saved on the site. New chat clears it.