Administration

Administration appears in the sidebar's "Your organisation" group and shows only the sections your organisation permissions allow. To prepare a new organisation step by step, follow Organisation setup.

SectionsCopy Anchor Link

Administration has these sections:

• Members (/app/members)
• Groups (/app/groups)
• Quota pools (/app/quota-pools)
• Analytics (/app/analytics)
• Providers (/app/providers)
• Models (/app/models)
• Organisation (/app/organization)

Which roles see which section is listed in Access control.

Instance-wide actions live in the separate Instance management area, outside the active organisation's Administration navigation.

MembersCopy Anchor Link

Members (/app/members) is a paginated data table with search over name and email, sorting, and table actions. It is scoped to the active organisation and acts on memberships. The row menu and the bulk action remove a member from the organisation, and the member's page changes the organisation role or removes the membership. The account itself survives and can be invited back. Banning an account and deleting it live under Manage instance > Users.

Primary actions:

• Invite member (/app/members/add)
• View invitations (/app/members/invitations)
• Remove from organisation in a row menu, and Remove selected from organisation for the selected rows

Invitation management has its own table page with export and actions. It lists every invitation of the active organisation, including those other admins and managers sent. The invitation form and the table are described under Invitations.

Edit member opens a member's page at /app/members/<user id>/edit. It shows:

• Member information with the name, email, status, email verification and dates.
• Effective access with every bot, block and asset the member can reach in this organisation, the role, and where the access comes from (a direct grant, a group, All members, an inherited bot or block, the organisation, or a Catalogue listing). Use it to find out why someone can or cannot see something. The table pages through the list and filters it by name, type and origin.
• Member management with Send password reset email, Send verification email while the address is unverified, the organisation role, and Remove from organisation.

Only an Admin can give the Admin role, change an Admin's role or remove an Admin, so a manager sees an Admin's page without the password reset, the role change or the removal. An organisation keeps at least one Admin who can sign in, so the last one cannot be demoted, removed, banned or deleted, and a banned Admin does not count. Your own page offers no password reset and no removal.

Removing a member from an organisation clears that organisation from the member's open sessions, and entering the app rechecks that the membership still exists.

Content always belongs to a Member, Manager or Admin of its organisation, so removing members and making a member a Viewer hand the bots, behaviours, repositories and Library assets they own there to one other member. The confirmation counts them and preselects you when you can own content there. Change picks another Member, Manager or Admin who is not banned and not part of the same change. Listings, additions in other organisations and grants stay as they are, and only the owner changes. Chats and chat uploads stay with the person. To split content between several people, receive it all and pass single items on with Hand over.

GroupsCopy Anchor Link

Groups (/app/groups) lists the groups with their member counts and a search. Custom groups can be created, edited and deleted, and system groups cannot be changed. A group opens at /app/groups/<group id>, where you add members and find the current ones in a searchable, paged table that marks each membership Explicit or Implicit. Access control covers the All members group and who may see what.

Quota poolsCopy Anchor Link

A quota pool bounds how much one provider, or one of its models, may be used in a period.

Quota pools (/app/quota-pools) lists the pools with their period, budget, consumed and remaining amounts. Add pool opens /app/quota-pools/add with these fields:

• Name, shown to members as the pool their usage is charged to, and an optional description.
• Provider, which the pool meters in the provider's metering mode, tokens or requests.
• Model scope, either All models from this provider or one model.
• Period, which is Weekly, Monthly or Yearly. The budget resets in UTC, weekly on Monday, monthly on the 1st and yearly on 1 January. Each period starts with the full budget, and unused budget does not roll over.
• Budget, the maximum tokens or requests per period.
• Priority, which decides between pools that could both apply, where the higher priority wins.
• Default pool, which marks the provider's default pool with a Default badge. One active pool per provider without a model scope can carry it.
• Active, which lets the pool take new usage.
• Eligible groups, whose people can use the pool. A pool with no eligible groups covers nobody, so choose All members to cover everyone.

Pools cannot be deleted. Deactivate pool, in the pool page's More actions menu, stops new usage, and usage already in progress still completes.

For each chat request, OrcAI looks at the active pools of the chat's provider whose eligible groups include the person. A pool scoped to the chat's model wins over pools covering all models. Between pools of the same scope, the higher priority wins, then the older pool. If no pool covers someone, they cannot chat with that provider's models.

Each call reserves its likely cost before the model runs and is charged what it used once it ends. With requests metering a reply reserves one request per step it may take, 25 in chat and 20 in the OrcAI Assistant, or what the pool has left when that is less. A reply starts while at least one request remains, gets as many steps as it reserved, and answers on the last of them, so it never uses more than the pool has. With tokens metering the pool needs the whole estimate of one request remaining: the prompt with the chat's instructions, tools and attachments, plus the chat's max tokens, or 16,384 tokens when the chat sets none, or the model's max output tokens when that is smaller. Titles, summaries, metadata and image descriptions reserve their own output limit, or else the model's max output tokens, or else 4,096 tokens. A reply that takes several steps can use more than it reserved, and it is charged all it used.

A call is charged the usage its provider reported for each finished request, also when it was stopped or failed after one. A request cut off in the middle reports nothing and is not charged, so usage is best-effort and not guaranteed to be exact. A web search through a search provider counts one request once it answers, also when it found nothing.

A call runs for at most 30 minutes, and OrcAI stores each finished request with its reservation while it runs. A reservation still open 35 minutes after it was made, such as one left by a server that stopped mid-reply, is settled by the workers within five minutes. It is charged the requests stored with it and shows as Settled, or it is released when none had finished.

View pool opens /app/quota-pools/<pool id>, with the current period's budget, consumed, reserved and remaining amounts, Adjust budget (raising the budget frees the remaining budget straight away), the eligible groups with their description, type and member count, and the usage events. The usage events table covers every period, newest first, and filters by user and event type: Reserved when a call starts, Settled with the amount it actually used, Released when a reservation is given back, and Failed when the provider failed. A stopped call that finished no request, or one that never ran, shows Released only. As in Analytics, people who have left the organisation and bots you cannot open stay unnamed, and the user search finds current members only. Edit pool opens /app/quota-pools/<pool id>/edit.

In chat, the usage indicator shows people the pool that covers their model as a share used, with its reset time, never amounts. People who can manage quota pools get a Manage pool link to the pool.

The OrcAI Assistant draws on the same pools as chat, through the organisation's Assistant model. Chat and conversation titles draw on the pools of the organisation's utility model when one is set. For a separate Assistant budget, create a pool for that model alone and assign it to All members, because a model-specific pool wins over the provider's pools. An exhausted pool blocks the Assistant like any chat.

AnalyticsCopy Anchor Link

Analytics (/app/analytics) shows how the organisation used its models over the last 7, 30 or 90 days, with 30 as the default. It reads the usage that quota pools metered, with tokens for every call and provider requests for every step, whatever the pool's metering mode. The two units are never added together.

Each call also carries an estimated cost, worked out when the call finished from the model's prices and its provider's currency at that moment. Changing a price later never changes the cost of past calls, and calls from before a price was set stay without a cost. Amounts in different currencies are never added together or converted, so a figure such as "€12.30 + $4.10" lists each currency. Admins enter the prices by hand, so the estimate may differ from the provider's invoice, and the page says so below the totals.

The page shows:

• The totals of tokens, input tokens, output tokens, requests, cache read and cache write tokens, and the estimated cost with one figure per currency. Below the cost, "Calls with no recorded cost" counts the calls whose model or search provider had no price.
• Tokens per day and Requests per day with one bar per UTC day, and Estimated cost per day with one series per currency.
• People with the usage per person, most tokens first. People who have left the organisation add up to one "Former members" row, without names or email addresses. Usage of deleted accounts adds up to one "Deleted users" row.
• Models with the usage per model and its provider. Usage recorded before analytics existed under a pool for all of a provider's models, and usage of a deleted model, shows as "Model not recorded".
• Bots with the usage in chats with a bot. Everything else, such as chats without a bot, the OrcAI Assistant and chat titles, counts as "Without a bot". Bots are recorded only since analytics were added, and usage of a deleted bot moves to "Without a bot". A bot shows by name only to an Admin who can open it, and all other bots add up to one "Other bots" row.
• Ratings by bot with how many replies were rated in the range, by the bot of their chat, with the count per rating and the up/down share, where up is "Very much" and "A bit". A rating counts for the organisation whose quota pool paid for the rated reply. Ratings of replies copied into a branch are not counted. Bots the Admin cannot open add up to "Other bots", as above. Message content is never shown.

The People, Models and Bots tables each have an Estimated cost column and stay sorted by tokens, then requests.

Only Admins see the page (view_analytics), and cost shows nowhere else, not even in a reply's usage. Deleting a provider deletes its usage records, so its usage leaves the page too.

ProvidersCopy Anchor Link

A provider is a service that runs models, or a search provider that serves web search, connected with its endpoint and key.

Providers (/app/providers) lists them with their status and metering mode. Add provider opens /app/providers/add with these fields:

• Kind, which is OpenAI-compatible for any endpoint that speaks the OpenAI API, OpenAI or Anthropic for their official APIs, or Tavily (web search) for a search provider. The kind is fixed once the provider exists, so to switch kind, create a new provider.
• Name, shown to members wherever the provider is named.
• Endpoint, the base URL of the endpoint including the version path, required for OpenAI-compatible providers. OpenAI and Anthropic providers have no endpoint field and use the official API. Tavily providers use Tavily's API while the field is blank.
• API key, stored encrypted. When editing, leave it blank to keep the stored key.
• Metering mode, whether quota pools count tokens or requests for this provider. Search providers always count requests, one per search.
• Currency, the currency the provider bills in, chosen from the ISO 4217 codes, USD for providers created before prices existed. Its models' prices are read in this currency. Changing it shows a warning that model prices are read in the new currency from now on, while past calls keep the currency they were recorded in.
• Reports cost (OpenRouter style), for OpenAI-compatible providers only, which uses each response's own cost, as OpenRouter sends it, instead of the model's prices. That cost is in USD, so the currency is locked to USD while this is on. A response without a cost falls back to the model's prices, and only a call whose every response reported one counts its cost as reported.
• Price per search, for search providers only, what one search costs for the estimated cost.
• Enabled, which lets the organisation use the provider's models.

A search provider has no models. It serves the organisation's web search once the Organisation page chooses it, and quota pools on it bound the searches. Its page warns while no active quota pool covers it.

A provider's page (/app/providers/<provider id>) shows its configuration, model counts and a searchable, paged table of its connected models, with Add model, Discover models and Edit provider (/app/providers/<provider id>/edit). There is no separate connection test, and Discover models, here or on the Models page, is a quick check that the endpoint and key work.

Discovery lists the models the provider offers from /models of the endpoint for OpenAI-compatible providers, /models of OpenAI's API for OpenAI providers, and Anthropic's Models API for Anthropic providers. It also takes each model's context window and max output tokens where the provider reports them. Anthropic's Models API does, and so do OpenAI-compatible servers such as OpenRouter, Together AI, vLLM and Groq. OpenAI's API reports neither.

When discovery fails, it says why: the provider rejected the API key, the endpoint does not offer a model list (add its models by hand), the provider could not be reached, or its answer could not be read.

With an Anthropic provider, OrcAI marks the system prompt and the messages of every request for Anthropic's prompt caching. The steps of an answer that uses tools, and a next answer within the cache's lifetime of about five minutes, read what was already sent at Anthropic's reduced cache rate. The Assistant declares all of its tools up front there and loads more of them through Anthropic's tool search, so loading tools or approving a plan keeps the cache. Claude models without tool search (Claude 3, and Claude Opus and Sonnet 4.1 and earlier) get loaded tools added to the request instead, like other providers. Claude's reasoning, when the model returns any, shows in the answer's work log as Thinking.

ModelsCopy Anchor Link

Models (/app/models) lists models separately from providers, with:

• Search by model or provider name
• Provider filter
• Capabilities filter, to find the models whose capabilities are not set yet
• Add model (/app/models/add)
• Discover models, which lists the models a provider offers, adds the new ones and fills in the token limits and capabilities a known model still lacks

A model has a name, an optional description, a provider, the Provider model ID the endpoint expects, its Capabilities, a Deprecated flag, and two token limits, empty when unknown:

• Context window, how many tokens the model reads at once. A chat reply shows how full it was, and a long chat is summarised once it fills 80% of the window after the reply's room, so a smaller value summarises earlier and keeps prompts cheaper. Empty means 128,000. A reply without max tokens asks for at most half of it.
• Max output tokens, how many tokens the model writes per reply at most. A chat's max tokens cannot be set higher, a reply asks for no more, and a limit below 16,384 lowers the pool reservation of a chat without max tokens.

Discovery sets a model's capabilities where its provider reports them:

• OpenAI-compatible providers take them from OpenRouter's model metadata (input and output modalities, and whether the model takes tools or reasoning). Other servers report none.
• Anthropic providers set Text and Tool calling, which every Claude model has, plus Image and Reasoning where the Models API reports them.
• OpenAI providers set none, because OpenAI's models API reports none.

Capabilities a provider does not report stay Not set, so set them yourself once on the models you want to use. Clearing every capability in the form makes them unknown again, and the next discovery fills them in where the provider reports them. Rediscovery never overwrites capabilities or a limit that are already set, so edited ones stay. A model opens at /app/models/<model id>, and Edit model at /app/models/<model id>/edit.

Model pricesCopy Anchor Link

Prices on a model say what its provider bills, in the provider's currency, for the estimated cost in analytics. Each is optional:

• Input, Cache read, Cache write and Output, per 1M tokens. Reasoning tokens count as output. An empty cache price uses the input price.
• Per request, added for every provider request.
• Per web search, for OpenAI and Anthropic models, for each search the model runs with its provider's own web search.

A model needs both an input and an output price to be priced. Without them its calls have no cost, unless its provider has Reports cost (OpenRouter style) on and every response of the call reported its cost. Once it is priced, an empty request or search price counts as zero. Prices apply to calls from the moment they are saved. They estimate what the provider bills and leave quota pools untouched, so pools keep counting tokens or requests. Long-context tiers, service tiers and longer-lived cache writes are not priced, so the estimate may differ from the invoice.

Instance managementCopy Anchor Link

Instance management is a separate application area under /instance, with its own sidebar and five pages. Instance admins can open it from the shared user menu or the organisation picker, including when they have no organisation memberships. /instance opens the Health page.

Health (/instance/health) shows whether the instance works. Each check is OK, a warning or failing, and is either required or recommended. PostgreSQL, SpiceDB, Qdrant and object storage are required. Embeddings is required too and passes once embeddings are set up and the active Qdrant collection exists with the model's dimensions. The page also requires a running workers process, and recommends email delivery over SMTP and a deployment without configuration warnings. A dependency that does not answer within three seconds counts as failing. Below the checks it lists the outbox (waiting and given-up emails, the latest error, and whether the latest email went out over SMTP or only to the worker log), the processes that run background jobs with their last heartbeat, the jobs per queue including dead letter queues, the OrcAI, PostgreSQL and Qdrant versions with the latest database migration, and the settings of the app process, including the stored embedding endpoint, model, dimensions and collection. Secrets only show whether they are set. The checks run when the page opens and on Check again, and nothing polls in the background. Send test email queues an email to your own address, at most one a minute, through the same outbox as every other email. Check again to see whether it was sent and how.

Settings (/instance/settings) has the Embeddings section, which sets the OpenAI-compatible endpoint and model that embed every organisation's repository assets and search queries. Save and test saves them only after a test embedding works. Embeddings describes the fields. Until they are set up, repository search is unavailable, and instance admins see a banner in the app that links to this page. Changing the model or the passage prefix embeds every asset in a repository again after a confirmation, and the Search index section counts the progress per organisation, with Retry failed assets, Prepare assets again and Re-embed everything, as described under Change the model.

Organisations (/instance/organizations) lists every organisation of the instance with its slug and member count, with search, creates one with Create organisation, and deletes one. Creating one asks for its Initial admin:

• Me, so you administer it.
• Existing user, so an account you pick administers it and you are not added. A banned account is refused.
• Invite by email, so you administer it for now and the address gets an admin invitation, both in one step. Once they have joined, leave the organisation from its page, because nothing removes you automatically.

An organisation's name opens its page (/instance/organizations/<organisation id>). It shows the organisation's readiness checklist read-only, the same checks its admins see on Home's Manage tab, and its members in a paged, sortable list, searchable over name and email. Each name links to the account's page. There an instance admin changes a member's role, removes a member (on your own row this reads Leave), Add existing user with a role, and Invite by email with a role. The organisation always keeps one admin who can sign in, so demoting or removing the last one is refused. Inviting someone who is already a member or already invited sends no email, as under Invitations, so change a member's role instead. Remove logo in the page's … menu deletes the organisation's logo. None of this gives an instance admin access to the organisation's bots, blocks and assets.

Before anything is deleted, the dialog shows an impact preview counting the memberships, pending invitations and groups that go with the organisation, plus the bots, behaviours, repositories and assets it is home to. It names those listed in the Catalogue, with the organisations that added them, since they lose them too. Delete is enabled only once the organisation's slug is typed back. The delete then removes all of them along with the organisation's chats, grants and authorisation relationships, and any session that had the organisation active loses it and lands back on the organisation selection page. Chats elsewhere that used a deleted bot become read-only.

Users (/instance/users) lists every account of the instance, searchable over name and email, with how many organisations each account belongs to, its instance role, and whether it uses two-factor authentication (2FA). An account's name opens its page (/instance/users/<user id>), which lists its organisations with its role and join date, each linking to the organisation's page, and carries the same actions in its … menu. From there you ban, unban or delete the account, and Remove picture deletes its profile picture. Reset 2FA helps someone who lost their authenticator app and backup codes. It removes their second factor, passkeys and trusted devices, signs them out everywhere and emails them, so they sign in with their password and set up 2FA again. Make instance admin and Remove instance admin change the account's instance role. None of these applies to your own account. Organisation admins do not see whether a member uses 2FA.

Deleting an account hands what it owns to one member per organisation, in every organisation where it owns a bot, behaviour, repository or Library asset, whether it is still a member there or not. The confirmation lists those organisations and preselects each one's earliest Admin who is not banned, and Change picks another. The account's chats and chat uploads are deleted with it.

Analytics (/instance/analytics) shows the instance's tokens and requests over the last 7, 30 or 90 days, per UTC day, and every organisation's totals with the number of people who used a model. Organisations without usage are listed with zeros. It shows the estimated cost per currency the same way as an organisation's Analytics.

The active organisation remains selected while the instance admin is in this area, but it does not scope or otherwise affect these pages. The Back to app action returns to /app, where the usual active-membership checks apply.

The instance adminCopy Anchor Link

Most administration is scoped to one organisation. A few actions concern the whole instance and belong to the instance admin, an account marked with the instance role in the database. The first account created when an instance is initialised is an instance admin.

Initialise an instanceCopy Anchor Link

A new instance has no accounts, so its sign-in and Create account pages lead to Initialise instance (/init). The form sets up the first organisation with its Organisation name and Organisation slug, and the first account with Admin name, Admin email, Password (at least 8 characters) and Confirm password. Privacy consent confirms that you agree to the instance's privacy policy and terms of use. Initialise instance creates the account as an instance admin and as Admin of the organisation, signs you in and opens the instance Settings page to set up embeddings. Set up later goes to the app instead, and uploads are not searchable until embeddings are set up. Since nothing is configured yet, Home opens on its Manage tab, where the Readiness checklist lists what the organisation still needs. Once any account or organisation exists, /init leads to the sign-in page.

An instance admin makes another account an instance admin, or takes the role away again, from the Users page. The account's owner gets an email either way. Nobody changes their own instance role or bans or deletes their own account, so an instance always keeps at least one instance admin, also when two instance admins demote, ban or delete each other at the same moment. With TWO_FACTOR_REQUIRED=admins, a new instance admin without two-factor authentication is sent to set it up the next time they open OrcAI.

Only an instance admin may:

• create an organisation,
• delete an organisation,
• ban and unban an account,
• reset an account's two-factor authentication,
• make an account an instance admin or remove the role,
• delete an account,
• remove an account's profile picture or an organisation's logo,
• list every organisation and every user of the instance,
• see every organisation's usage.

The instance role grants nothing inside an organisation. An instance admin still needs a membership to enter an organisation or manage its members, and an organisation admin manages memberships only.

OrganisationCopy Anchor Link

The Organisation page (/app/organization) under Administration edits the active organisation's name, slug, logo and description and shows its ID and creation date. Only admins of the organisation see it.

About the organisation is the description everyone on the instance reads on the organisation's page (see Catalogue). It takes headings, bold and italic text, lists, quotes and links to web, email (mailto:) and phone (tel:) addresses, up to about 20,000 characters of stored content. Put contact details into the text. Emptying it removes it.

Change logo uploads a square logo the same way as a profile picture: pick or drop one PNG, JPEG, WebP or GIF of up to 10 MB, crop it, and the browser stores a 256 by 256 pixel WebP. Remove logo deletes it, and the organisation's initials show instead. The logo shows on the organisation picker, invitations, next to the organisation in the user menu and in instance management. The sidebar keeps the OrcAI logo, and emails show no logo. Organisation admins cannot remove members' pictures, which belong to the account across organisations, but an instance admin can.

To work on another organisation, switch to it first with Switch organisation in the user menu, which opens /select-organization. Creating and deleting an organisation are instance actions and live under Manage instance > Organisations.

Assistant modelCopy Anchor Link

The Assistant model picker on the same page chooses the model the OrcAI Assistant answers with. It is the chat model picker, limited to models with tool calling that are not deprecated and run on an enabled provider. Clear removes the choice.

While no model is set, the Assistant is unavailable to everyone in the organisation. Deleting the model, or its provider, clears the setting. If the chosen model later becomes unusable, because it is deprecated, loses tool calling or its provider is disabled, the setting stays, the page warns with the reason, and the Assistant's replies fail until another model is chosen.

Default chat modelCopy Anchor Link

The Default chat model picker chooses the model new chats start on. It offers the same models as the Assistant model picker. Clear removes the choice.

A bot's required model always decides, and a model picked in the composer comes next. Otherwise OrcAI preselects the first usable model that an active quota pool covers for the person starting the chat, in this order: the model the bot recommends, the default chat model, then the organisation's newest chat model. If no pool covers any of them, the first usable one is preselected and the composer says that no pool covers it. A chat created without a model, for example through the API, follows the same order. If its bot recommends a model, the chat stores the model the order picks for its creator. Otherwise it answers each message with the model the order picks for the person sending it.

While no default is set, the order goes straight from the bot's recommendation to the organisation's newest chat model. If the chosen model becomes unusable, because it is deprecated, loses tool calling or its provider is disabled, the setting stays, the page warns with the reason, and new chats skip it until it is usable again. Deleting the model, or its provider, clears the setting.

Utility modelCopy Anchor Link

The Utility model picker chooses the model that writes the titles of new chats and Assistant conversations, describes images in the Library and generates asset metadata, so a small, inexpensive model can do this work instead of each chat's own model. It offers text models that are not deprecated and run on an enabled provider, without requiring tool calling. Clear removes the choice.

Title calls draw on the quota pools of the utility model, like any chat on that model. While no utility model is set, each chat's title is written by its own model. If the chosen model becomes unusable, the page warns and titles fall back to each chat's own model until another is chosen. Deleting the model, or its provider, clears the setting.

Images are described once, when an asset is prepared: image files and the figures inside files. This needs a utility model with the image capability. Without one, image files fail to prepare and figures inside files are skipped. Each description draws on the quota pools of the utility model, attributed to the person who uploaded the file. When the pool is used up, the file fails to prepare and can be retried once the pool resets. When no active pool covers the utility model for the uploader, create or adjust one before retrying.

The utility model also generates asset metadata, automatically for prepared Library uploads and on request from the metadata editor, the Library's Assets view or the Assistant. Each run draws on the quota pools of the utility model, attributed to the uploader for automatic runs and to the person who asked otherwise. Without a utility model, no metadata is generated.

Web searchCopy Anchor Link

The Web search setting decides how bots with the Web search feature search the web. It is off by default.

• Off, so no bot searches the web.
• Model's own search (Anthropic and OpenAI models), so Anthropic and OpenAI models use their provider's own web search tool, and an Anthropic model searches at most five times per request. Models on other kinds of provider get no search. The provider charges each search on its own bill, which OrcAI does not meter, and the search results the model reads count as tokens like any other input.
• Search provider: name, so every model searches through one of the organisation's search providers, such as Tavily. Each search reserves one request on the provider's quota pools and is charged once it answered, and a failed search is not charged. When no pool covers the person, or the pool is used up, the model is told that web search is unavailable and answers without it.

Choosing a search provider that no active quota pool covers shows a warning below the setting, and on the provider's page, with a link to add a quota pool.

A bot's allowed domains apply in both modes: OrcAI passes them to the search and drops any search provider result whose host is not one of them or below one. A bot chat whose web search cannot run, because the setting is off or the model has no search of its own, shows a notice above the composer and answers without searching. Deleting the chosen search provider turns web search off.

Operational guidanceCopy Anchor Link

• Configure providers and models before inviting many members.
• Keep quota pools aligned with model pricing and expected usage.
• Use groups and grants to avoid manually maintaining per-user grants.