Admin dashboard Settings

Models & provider

Models & provider sets the AI every member’s assistant runs on: the provider, your organization’s key for it, and the model and effort for each kind of call. What the assistants work from goes to that provider.

Who changes itAdmins and owners
Applies toEvery member’s assistant
KeyYour organization’s own
Takes effectEach member’s next turn
In the admin dashboard: SettingsModels & provider Open the dashboard
On this page

Provider and key#

Every member’s assistant runs on the provider you choose here, and so does what it does in the background: its wiki, its retrospectives, its jobs and the line that says what the member is doing. Each kind of that work runs on its own model of that provider (below). Members can’t change them.

FeatureDescriptionDefault
Provider

The AI provider every member’s assistant runs on: Anthropic, OpenAI, OpenRouter, an OpenAI-compatible server, or an Ollama server. Anthropic, OpenAI and OpenRouter run on your organization’s own key, so you can’t save one of them as the provider without its key. An OpenAI-compatible server can’t be saved without its address. A server of your own can’t be saved as the provider until every kind of call has one of its models.

Anthropic
API key

Your organization’s key for the chosen provider. The field is named for it: Anthropic API key (from console.anthropic.com), OpenAI API key (from platform.openai.com) or OpenRouter API key (from openrouter.ai). For an OpenAI-compatible or Ollama server it’s just API key, needed only if the server checks one, as Ollama’s cloud does.

Each provider keeps its own key. A key is encrypted before it’s stored, and once saved it’s never shown again: the field only says that one is saved. Type a new one to replace it, or choose Clear the saved key. The key of the provider you run on can’t be cleared while that provider needs one.

Not set
Base URL

The address of your OpenAI-compatible server, shown when that’s the provider. It must be a public https address: an address in a private or local network is refused, and so is one with a user name or password in it. It’s checked when you save.

None
Ollama address

The address of your Ollama server, shown when that’s the provider: https://ollama.com for Ollama’s cloud, or a public https address of your own. The same rules apply as for Base URL.

None

Your organization starts on Anthropic with no key, so assistants can’t answer until you add your key or set up another provider. Members’ messages wait, and are answered once you do.

The providers#

  • Anthropic: Anthropic’s API, on your organization’s key.
  • OpenAI: OpenAI’s API, on your organization’s key.
  • OpenRouter: OpenRouter, on your organization’s key. It passes each request on to the company behind the model you pick, so that company receives it too.
  • OpenAI-compatible server: any server that speaks OpenAI’s API, at a public https address you give.
  • Ollama server: an Ollama server at a public https address you give, or Ollama’s own cloud at https://ollama.com, with your key.

A model for each kind of call#

Each kind of work the assistants do runs on its own model of your provider, at its own effort. There is no main model. At first nothing is chosen, so each kind runs on the model and the effort we recommend for it (see the models we recommend). To choose another model, select one of the provider’s models in the kind’s list. The list puts the models we list for your provider first, under Recommended, and then the other models that your provider lists, under All models. To type a model that the list does not show, select Custom… and type its name. Under the model, select the kind’s effort. Use recommended puts one kind back on the recommended model and effort, and Use recommended for all kinds puts every kind back.

An Ollama or OpenAI-compatible server of your own has no recommended models, because only the server knows which models it serves. Each kind’s list shows the models the server serves, as the server listed them last. Select Refresh list to ask the server again, for example after you add a model to it. Select a model for one kind, then select Apply to all kinds to give every kind the same model. You can’t save the server as your provider until every kind has a model.

Each provider keeps its own choices: choosing another provider shows that provider’s, switching back doesn’t lose them, and a switch never leaves a kind on a model the new provider doesn’t serve. The admin dashboard’s Model usage shows the models each kind ran on.

FeatureDescriptionDefault
Ambient turns

The model for the assistant’s turns over what members’ desktops hear and record outside meetings, when no message from them waits for an answer. Most of the assistant’s turns are these. When you choose no model, the recommended model runs.

The recommended model
Chat

The model for the assistant’s answers to a member’s own messages outside meetings. When you choose no model, the recommended model runs.

The recommended model
Meetings

The model for the assistant’s turns while a meeting is live, its answers to messages during the meeting among them, the meeting’s recap, and its answers to questions about the meeting afterwards. When you choose no model, the recommended model runs.

The recommended model
Doing summarizer

The model that reads the activity members’ desktops record, and writes the short line that says what each member is doing. Their assistant reads that line, and their timeline keeps it. When you choose no model, the recommended model runs.

The recommended model
Work jobs

The model for members’ jobs: the work they accept from their assistant’s offers or ask for in chat. The same model checks how each job went. When you choose no model, the recommended model runs.

The recommended model
Work scout

The model for the scout that looks for work each member’s assistant could offer them, every four hours. When you choose no model, the recommended model runs.

The recommended model
Wiki upkeep

The model that folds each member’s journal and meeting recaps into their wiki, and keeps each team’s and the organization’s wiki from the work of their members. When you choose no model, the recommended model runs.

The recommended model
Page descriptions

The model that writes the line each wiki page is listed by. When you choose no model, the recommended model runs.

The recommended model
Retrospectives

The model that writes the retrospectives: each member’s, and each team’s and the organization’s. When you choose no model, the recommended model runs.

The recommended model
Effort

How much each kind’s model thinks before it answers. Each kind has its own list, under its model. The list uses your provider’s own words: Effort on Anthropic (low, medium, high, xhigh, max), Reasoning effort on OpenAI and OpenRouter, and Thinking on an Ollama server (off, on, low, medium, high). It shows only the levels that the kind’s model takes. A model that takes none, for example Claude Haiku 4.5, shows no list. Provider default sends no level, so your provider’s own default applies.

When you choose no level, the recommended effort applies if the kind’s model takes it. If it does not, no level is sent. When you choose another model for a kind, a level that the new model does not take is cleared.

The recommended effort

These are the models each kind runs on when you choose none, at the effort shown. Provider default means that no level is sent. Each kind gets a model that fits its work: a larger model where the answer needs judgment, and a small model where the work is short and simple. No kind runs on a flagship model (Claude Fable 5.1, GPT-6 Astra) unless you choose it. Your provider sets and changes the price of each model, so check your provider’s current pricing to understand what each kind of call costs you.

How often each kind runs, for one member:

  • Ambient turns: at most once every 2 minutes outside meetings, at the default Assistant pace.
  • Chat: once for each message a member sends.
  • Meetings: as often as every 20 seconds while a meeting is live, and once for its recap.
  • Doing summarizer: often, while the member’s desktop records activity.
  • Work jobs: once for each job, and once more to check it.
  • Work scout: every four hours.
  • Wiki upkeep and page descriptions: only when something changed.
  • Retrospectives: once a day.
Kind of callModelEffort
Ambient turnsClaude Opus 5.5claude-opus-5-5medium
Ambient turns are most of the assistant’s turns, and in each one it decides by itself what in the member’s day to keep. That takes good judgment, so Opus 5.5, a large model, runs them at medium effort, its own default.
ChatClaude Opus 5.5claude-opus-5-5medium
Chat runs once for each message a member sends, and the answer is what the member asked for. Opus 5.5 is a large model, and at medium effort, its own default, it gives a strong answer.
MeetingsClaude Sonnet 5.5claude-sonnet-5-5medium
Live meeting turns come often and must keep up with the talk, and the recap needs good judgment. Sonnet 5.5 is a mid-sized model that balances speed and judgment, and medium effort suits both.
Doing summarizerClaude Haiku 5.5claude-haiku-5-5low
The doing summarizer runs often, reads the member’s latest activity and writes one short line. Haiku 5.5 is a small model, and low effort is enough for work this short and simple.
Work jobsClaude Opus 5.5claude-opus-5-5high
A work job is real work with tools that a member asked for, and a check reads each result once. Jobs run rarely, so Opus 5.5 at high effort can take its time on each one.
Work scoutClaude Sonnet 5.5claude-sonnet-5-5low
The scout reads a large part of the member’s recent record and must judge which work to offer. Sonnet 5.5 is a mid-sized model with room for that much reading, and low effort is enough for a short list.
Wiki upkeepClaude Opus 5.5claude-opus-5-5medium
Wiki upkeep rewrites what the wiki knows, and an error stays in the wiki until someone finds it. Opus 5.5 at medium effort gives careful edits, and the upkeep runs only when something changed.
Page descriptionsClaude Haiku 5.5claude-haiku-5-5low
A page description is one line for each wiki page. Haiku 5.5 is a small model, and low effort is enough for one line.
RetrospectivesClaude Opus 5.5claude-opus-5-5medium
Retrospectives run once a day and are written for people to read. Opus 5.5 at medium effort gives careful writing.
Kind of callModelReasoning effort
Ambient turnsGPT-6.1 Solgpt-6.1-solmedium
Ambient turns are most of the assistant’s turns, and in each one it decides by itself what in the member’s day to keep. That takes good judgment, so GPT-6.1 Sol, a large model, runs them at medium effort, its own default.
ChatGPT-6.1 Solgpt-6.1-solmedium
Chat runs once for each message a member sends, and the answer is what the member asked for. GPT-6.1 Sol is a large model, and at medium effort, its own default, it gives a strong answer.
MeetingsGPT-6.1 Solgpt-6.1-solmedium
Live meeting turns come often and must keep up with the talk, and the recap needs good judgment. GPT-6.1 Sol is a large model, and medium effort, its own default, suits both.
Doing summarizerGPT-6 Lunagpt-6-lunalow
The doing summarizer runs often, reads the member’s latest activity and writes one short line. GPT-6 Luna is a small model, and low effort is enough for work this short and simple.
Work jobsGPT-6.1 Solgpt-6.1-solhigh
A work job is real work with tools that a member asked for, and a check reads each result once. Jobs run rarely, so GPT-6.1 Sol at high effort can take its time on each one.
Work scoutGPT-6.1 Solgpt-6.1-sollow
The scout reads a large part of the member’s recent record and must judge which work to offer. GPT-6.1 Sol is a large model with room for that much reading, and low effort is enough for a short list.
Wiki upkeepGPT-6.1 Solgpt-6.1-solmedium
Wiki upkeep rewrites what the wiki knows, and an error stays in the wiki until someone finds it. GPT-6.1 Sol at medium effort gives careful edits, and the upkeep runs only when something changed.
Page descriptionsGPT-6 Lunagpt-6-lunalow
A page description is one line for each wiki page. GPT-6 Luna is a small model, and low effort is enough for one line.
RetrospectivesGPT-6.1 Solgpt-6.1-solmedium
Retrospectives run once a day and are written for people to read. GPT-6.1 Sol at medium effort gives careful writing.
Kind of callModelReasoning effort
Ambient turnsClaude Opus 5.5anthropic/claude-opus-5.5Provider default
Ambient turns are most of the assistant’s turns, and in each one it decides by itself what in the member’s day to keep. That takes good judgment, so Opus 5.5, a large model, runs them.
ChatClaude Opus 5.5anthropic/claude-opus-5.5Provider default
Chat runs once for each message a member sends, and the answer is what the member asked for. Opus 5.5 is a large model that gives a strong answer.
MeetingsClaude Sonnet 5.5anthropic/claude-sonnet-5.5Provider default
Live meeting turns must keep up with the talk, and the recap needs good judgment. Sonnet 5.5 is a mid-sized model that balances speed and judgment.
Doing summarizerClaude Haiku 5.5anthropic/claude-haiku-5.5Provider default
The doing summarizer runs often and writes one short line each time. Haiku 5.5 is a small model, and that is enough for work this short and simple.
Work jobsClaude Opus 5.5anthropic/claude-opus-5.5Provider default
A work job is real work with tools that a member asked for. Jobs run rarely, so Opus 5.5 can give careful work on each one.
Work scoutClaude Sonnet 5.5anthropic/claude-sonnet-5.5Provider default
The scout reads a large part of the member’s recent record and must judge which work to offer. Sonnet 5.5 is a mid-sized model with room for that much reading.
Wiki upkeepClaude Opus 5.5anthropic/claude-opus-5.5Provider default
Wiki upkeep rewrites what the wiki knows, and an error stays in the wiki until someone finds it. Opus 5.5 gives careful edits, and the upkeep runs only when something changed.
Page descriptionsClaude Haiku 5.5anthropic/claude-haiku-5.5Provider default
A page description is one line for each wiki page. Haiku 5.5 is a small model, and that is enough for one line.
RetrospectivesClaude Opus 5.5anthropic/claude-opus-5.5Provider default
Retrospectives run once a day and are written for people to read. Opus 5.5 gives careful writing.

The recommendations send no effort to OpenRouter, because OpenRouter does not say how it passes an effort on to Claude 5.5 models. So OpenRouter’s own default applies: its catalog gives high for Opus 5.5 and Sonnet 5.5, and medium for Haiku 5.5. A level that you choose for a kind is sent.

No defaults, because only your server knows which models it serves. Select a model for one kind, then select Apply to all kinds. These servers have no standard field for an effort, so an effort goes only to a model for which Ambient Assist AI knows the field.

No defaults, because only your Ollama server knows which models it serves. Select a model for one kind, then select Apply to all kinds. Thinking is offered only for a model whose levels Ambient Assist AI knows.

What your provider receives#

Your provider receives what the assistants work from, as text: what members’ desktops hear, members’ messages to their assistant, their wiki pages, and the activity desktops record that you allow. When someone with the Team or Org role asks their assistant about the people they lead, those people’s records go to it too.

The words people read and copy on their desktops reach your provider only if you enable Send content to the model under Desktop awareness.

  • Desktop awareness: what desktops record besides speech, and whether its text reaches your provider.
  • Intelligence & work: the retrospectives and jobs that run on these models.
  • Assistant pace: how often each member’s assistant is asked, which sets how often ambient turns run.
  • Model usage: the requests and tokens of each kind of call, and the models they ran on.
  • Roles and permissions: who can change these settings, and whom a lead’s assistant can ask about.
Esc
↑ ↓ to moveEnter to open