The AI features of CRM & Studio are five agent runners in the drop-core-generative package — one tool loop that runs on Gemini or Claude, five system prompts and tool sets. Each runs on behalf of a signed-in user, writes through the normal API with CRM & Studio's credential, and is metered per turn (see Credits & metering).
| Runner | Where in CRM & Studio | What it does |
|---|---|---|
| web-builder | Document editor (Assistant pane) | Edits one page or document: content, sections, element placements, SEO — within the store's schema. |
| design-system | Studio › Design | Proposes and commits design tokens (colors, type, radius) with a live preview; interactive pickers pause the agent for the user's choice. |
| schema-builder | Settings › Schemas | Creates and evolves schemas — field definitions, enums, references — and registers collections. |
| template | Settings › Templates | Writes versioned Handlebars template bundles for a schema, with the rules for bindings, CSS scoping and load-more. |
| code | Studio › Code Studio | The unified editor over the whole store's code-bearing artifacts: design CSS, templates, store templates, setup — with the open file as context. |
All five share the canonical rule fragments (rules/*.md, the Rules pages) — they are inlined into the system prompts at build time and served to external agents as MCP resources, so an MCP client follows the same conventions CRM & Studio's own agents do.
Models#
The model menu in the Assistant offers Gemini (Flash — the default, Pro, Argon) and Claude (Sonnet 5.5, Opus 5.5, Fable 5.1). Every model has its own price: each row of the menu shows its rate in credits per 1M tokens, and the full rate card is on Settings › Plan › Pricing. A thread can change model at any time — the conversation carries over, only the previous model's private reasoning does not. The design-system runner always uses Gemini Pro. Image generation is Gemini-only (Nanobanana Lite / Pro), whatever text model is picked.
For hosts embedding the runners, each runner takes apiKey (Gemini — always required, for image generation and the fetch classifier) and an optional anthropic: {apiKey, workspaceId} (required when the turn's model is a Claude one). Both providers get the same system prompt and tool set.
Prompt stability#
A runner's system prompt depends only on the store id, its locales, the design version and a few mode flags. Everything per turn (the open file, the selection, the reader's message) travels in the user content or behind a read tool. That keeps the model's cached prefix — history included — valid across turns, which is what makes long sessions affordable. The package README's prompt budget section describes the check.