Model, payment and voice
The model is the agent’s intelligence engine. The voice determines how its spoken replies and voice notes sound.
Choose the source of intelligence
Section titled “Choose the source of intelligence”Open Agent → Settings → Intelligence. The page shows each use of the agent on one line:
- Main model: conversations, scheduled tasks and channels;
- Fallback model (optional): takes over if the main model fails;
- Memory: extracts and consolidates memories, with its own fallback;
- Voice calls: the reasoning model for spoken conversations, with its connection and its reasoning level;
- Auxiliary model: small background tasks (conversation titles, dictation correction). A fast, light model is enough.
Edit opens the list of models available for that use: Diadems models, then your subscriptions and your keys. Choosing a model from your OpenAI subscription, your Claude subscription or your key for the main model moves conversations onto that connection.
Reasoning is set directly on the line for each use, with only the levels its model offers. Conversation reasoning uses “High” by default. You can save another level for the agent; choices made for a single session still take priority. Memory reasoning is independent: the OpenAI subscription starts on “Low”, while other connections keep the model’s default until you choose a level. The fallback model takes the main model’s reasoning, adapted to what it accepts: its line says so. The auxiliary model has no reasoning level: conversation titles and dictation correction always run without reasoning, to stay fast. Your existing settings are kept. A higher effort can increase duration and consumption. The levels offered depend on the model; the provider may adapt or ignore this setting depending on its capabilities.
Four connections can be used:
- Diadems credits: a catalogue you can use without an external key, with models processed in France, in the European Union or with no EU guarantee;
- My API key: a direct call to the provider you choose with your own key. Use my API key, at the bottom of the main model list, asks for the provider, the key and the exact model identifier. Once the key is stored (a Mistral subscription connected in the Vault is one), its models appear in the list under My key;
- My OpenAI subscription: use of your ChatGPT/Codex account once connected from this agent’s Vault;
- My Claude subscription: experimental use of your Claude Pro or Max account through Claude Code, once connected in the Vault.
The Diadems catalogue shows the guaranteed processing location, the level of intelligence and the input and output prices. The France or EU flag matches the model’s inference location.
Your choices stay pending: a bar at the bottom of the page shows the number of changes and what will restart. Apply, then Confirm and apply, save them in one go; Cancel discards them. While changes are pending, the page asks you to apply or cancel them before you leave. The agent keeps its capsule and its data; messaging reloads for a few seconds, as does the memory service if its model changes, and a run in progress may be interrupted.
Only managers change the connection or set the memory and the auxiliary model. Other members with access to settings choose among the models of the connection already in place.
Use your Claude subscription
Section titled “Use your Claude subscription”In Vault → Connected accounts → Claude, select Connect Claude, open the link, then paste the code provided by Claude. The attempt expires after ten minutes and can be cancelled. Then choose My Claude subscription in the Edit list of the Main model in Settings → Intelligence, and apply. Sessions can also choose this connection.
This integration is experimental. Quotas are shared with the account’s other uses. Some models and overages may use usage credits billed by Anthropic; the selector shows the account’s indications when they are available. Check your Claude settings to limit overages. The portal does not show a Claude quota counter.
An Anthropic API key or another Claude backend registered for this agent can prevent this connection: the Vault explains the conflict. The memory model keeps its independent connection. No paid fallback model is added when you choose Claude; the fallback already chosen stays shown on its line.
If the Claude environment is not yet enabled on the capsule, the connection is shown as unavailable and must be prepared by Diadems.
Track your OpenAI subscription
Section titled “Track your OpenAI subscription”When an OpenAI account is connected to the agent, Settings → Intelligence shows the quotas reported by OpenAI: percentage remaining, next reset and date of the last reading. These limits apply to the whole connected account, including its other uses.
The reading refreshes every minute while the page is visible. Refresh lets you check again. Unavailable information does not mean the quota is used up; if refreshing fails, the previous reading stays, dated and flagged. This check triggers no generation and buys nothing on the subscription.
Manage my subscription opens Vault → OpenAI directly, to view or change this agent’s connection.
If the AI connection can no longer answer
Section titled “If the AI connection can no longer answer”The chat explains the known refusals: OpenAI subscription limit, Diadems credits used up, your provider’s quota, a connection to renew or a temporary incident. When OpenAI gives a renewal date, it is shown.
When a fallback model is available, the card offers Use followed by its name, with its connection and payment method. This button changes the model of the open session and resends your last request once. Actions already carried out may be repeated. Nothing is resent if the change is not confirmed.
Choose another model opens Settings → Intelligence. Members without this right must ask a manager. If the Diadems balance is used up, a manager can choose Top up Diadems credits, or Check the automatic top-up if it is already turned on, which opens Budget. Another model from the same exhausted subscription is not offered as a fix for its quota.
While settings that require a restart are being applied, all members viewing the agent see a waiting screen. It disappears once availability is confirmed; the thread and drafts stay in place. If the check fails or takes too long, you can check the status or go back to the list of agents. An interrupted conversation never restarts by itself.
Understand payment
Section titled “Understand payment”With Diadems models, usage is deducted from the credit offered with the first payment, then from the rechargeable reserve, at the provider’s price; the agent’s budget lets you track the share consumed.
With a personal API key or an OpenAI subscription, text does not use the Diadems AI budget. The provider or subscription concerned applies its own limits and terms.
Plan a fallback model
Section titled “Plan a fallback model”Under the model choice, Fallback model sets a second connection used only if the first fails during a reply: OpenAI subscription at its limit, provider error, rate limit. The agent finishes its reply with this model, then goes back to the main model by itself on the next turn.
The fallback can be a Diadems model, your OpenAI subscription or a personal key already saved in the Vault, as long as it differs from the main model. A Diadems fallback model is deducted from the free credit, then from the reserve; the account’s automatic top-up, if you have turned it on, tops it up as for any other use. The setting is applied with Confirm and apply, like the rest of the settings.
Call your agent
Section titled “Call your agent”In an available session, select Call next to the model. The microphone stays open: speak normally, wait for the spoken reply, then carry on without holding the mic button. Mute temporarily stops listening and Hang up ends the call. You can also say “hang up”.
This is the same session as the chat, not a separate voice assistant. Each request is transcribed, handed to Hermes with the model chosen in Settings → Intelligence → Voice calls, then the reply is read aloud with the chosen voice. When you call, the agent simply says “Yes?” as soon as the line is ready: this fixed phrase is played without calling a reasoning model. The agent keeps its personality, history, memory, skills, tools and approval requests. Requests handed to Hermes and its replies stay in the conversation. The call details also show the greeting exchanges during the call.
You can interrupt a reply by speaking again. Diadems stops playback and, if needed, the work in progress before passing on your new sentence.
The model for calls is chosen in Settings → Intelligence → Voice calls. Your choice is saved directly and does not restart the agent. With no specific choice, it follows the conversation model. A high reasoning effort can lengthen the delay before the reply. Memory keeps its own setting.
To add an authorised team member, ask the agent to invite them to the call. Each participant speaks with their own microphone and the agent’s replies are shared. Each person keeps their own permissions; the system does not identify a person by voice biometrics.
Choose a voice
Section titled “Choose a voice”- Open Agent → Settings.
- Open Voice.
- Click Listen to compare the samples.
- Select the voice you want.
You can choose from seven native Mistral voices compatible with context adaptation: Marie, Manon, Léa and Camille, French voices, and David, Fred and Hugo, British male voices that also speak French. Marie remains the default voice. Former cloned voices and fixed-emotion variants are replaced by Marie; you can choose another voice.
The new choice applies to the next calls, to the chat’s Listen button and to voice notes sent through compatible channels. During a call, the synthesis receives an excerpt of the sound context so it can adapt its tone. This context is temporary, kept in memory during the call, with no audio recording to disk by Diadems.
If no sound starts, check the volume and the browser’s audio playback permission, then try again with a sample.
Navigation and memory model
Section titled “Navigation and memory model”In Settings, open Intelligence to set the connection and the models together. The page opens with a breadcrumb; a confirmation always comes before changes that reload the agent.
The memory model is set on the Memory line of the same page, with its own fallback model, and is applied with the same Confirm and apply: a single reload covers all the settings.
Auxiliary model
Section titled “Auxiliary model”The agent writes small supporting texts itself: the title of each conversation and the correction of your dictated messages. Settings → Options lets you turn off dictation correction for this agent.
When its answer leaves you a real choice of what comes next, the agent itself offers two buttons under its latest reply: the follow-up it recommends and an option that goes further and takes more work, then Something else… to write your own. A button sends its message as if you had typed it; nothing leaves without your click. These buttons are written with the answer by the main model: they cost no extra call.
By default, these tasks use the main model. The Auxiliary model line in Settings → Intelligence lets you choose a smaller, faster model among the Diadems models or your connections. It is billed on its own connection, like the conversation. The summary of long conversations and image analysis keep the conversation model.