Organization
Defaults that apply to everyone, and the switches that decide what members may do at all.
Organization in the console. Two kinds of thing: defaults that shape every answer, and switches that turn features on or off for the whole organisation.
Name and context
Organization name is what appears around the product.
Organization context is background about your company that the assistant can draw on: what you do, what you sell, the vocabulary you use.
Worth five minutes. "We are a B2B software company selling compliance tooling to mid-sized banks. Our product is called Ledgerline. Our customers are compliance officers, not developers." That single paragraph improves every answer every member gets, because it stops the assistant guessing your industry from context clues.
System instructions
Default instructions applied to all AI responses in your organisation.
Use them for things that are true for everybody: house tone, the language to answer in, standing prohibitions.
Keep these short and universal.
Everything you put here is in force for every message from every member, and it competes for attention with whatever the user actually asked. A long list of special cases makes the important rules weaker, not stronger. Anything that applies to one team or one job belongs in an agent or a skill, which is what those are for.
Colour palette
The colours used in charts, presentations and visualisations. Set your brand colours once and everything generated comes out on-brand, with presets if you have none to hand.
Presentation template
Set your brand deck once and generated presentations come out on-brand: layouts, fonts, colours.
Two ways to give it one, and they are alternatives rather than steps.
Upload a .pptx, up to 50 MB. After upload you are told what was found (layouts, slides, images) and warned where something will not convert perfectly. Charts, SmartArt and embedded objects come through as images rather than editable elements, which is worth knowing before a client sees the deck.
Or point at a Google Slides deck, by pasting its link.
Pointing at a Slides deck does two jobs at once, which is the part worth understanding:
- Members with a linked Google account get decks built directly in Slides, on that deck's own layouts. Nothing is exported, so no size limit applies.
- Everyone else gets a PowerPoint file from the same design.
So one field covers both populations. You do not need to also upload a .pptx as a fallback.
The deck is not kept in sync. We keep a copy as your branding, so point at it again after you change it.
Turn off stops decks being built in Slides, and everyone gets PowerPoint files again.
Template analysis works best with a vision-capable chat model configured. Without one, layout detection is weaker. See Models.
Web search
Turn web search on and choose a provider, with an API key where the provider needs one. There are three ways to get search, and they differ more than the list suggests.
The three options
| Key needed | Cost | What comes back | |
|---|---|---|---|
| Tavily | Yes | Paid | Results, a written answer, relevance scores and images |
| DuckDuckGo | No | Free | Results and page content |
| The model's own search | No | Billed by your model provider | Whatever that provider returns |
Tavily is the richest. It is the only one that returns a written answer alongside the results, and the only one that scores each result for relevance, which is what lets the assistant tell a strong source from a passing mention. It is also the only paid-for choice with a bill of its own.
DuckDuckGo returns links, and pages are read separately to get at the text. There are no relevance scores, so results arrive in the order the engine gave them. DuckDuckGo needs no account at all, which makes it the quickest way to try the feature, though a free service rate-limits heavy use, and a search that hits the limit comes back empty rather than as an error.
The model's own search needs no provider and no key. Models that can do this are marked with Built-in web search in your model list, which today covers the current Anthropic, OpenAI, Google, xAI and Groq models. It is billed by whichever provider runs the model, so it lands in your ordinary usage rather than on a separate invoice.
Which one is used
Prefer the model's own search decides this, and it is the setting to understand.
On, a model that can search does so itself, and your configured provider is the fallback for models that cannot. Off, everything goes through the configured provider, whatever the model is capable of.
Leave it on. It keeps search on the same bill as the model that ran it, saves you a second vendor for most traffic, and gives each model the search it was built around. The reason to turn it off is wanting every search to go through one vendor, such as a search provider you already audit.
One exception applies whatever you choose: a few Google models cannot combine their own search with the assistant's other tools, so those fall back to the configured provider on their own.
Choosing
Most organisations should turn Prefer the model's own search on and configure Tavily behind it, which gives the better search to models that have one and a capable fallback to the rest. DuckDuckGo is the right choice for trying the feature or for light use.
On Asteria Cloud credits, search is set up for you and drawn from the same balance as everything else, so there is no separate account to open and no key to manage. That is also why the provider and its key are fixed here.
Feature switches
Three, each turning a capability on or off for everybody.
Allow users to add personal MCP servers. Off means members cannot connect their own tool servers. Off is the safer default: a personal server is a tool endpoint you have not reviewed. See MCP servers.
Agents marketplace. Lets members publish and browse agents internally. Disabling hides the Browse tab and blocks new publishing, while existing library entries keep working, so turning it off is not a rug-pull. Comes with a link to the moderation queue.
Allow users to create projects. When on, anyone can create a project, and every new project starts with the default budget cap you set here. API usage stops at that cap until an admin raises it. That cap is what makes this switch safe to leave on: it bounds the blast radius of somebody creating a project and pointing production traffic at it.
Admin API keys
Keys prefixed sk-ast-adm- for the organisation management API. They belong to no project.
An admin key can create users, mint project keys and read organisation-wide usage. It is not a bigger project key, it is a different thing, and it does not belong in an application. See Projects.