Create groups, manage members and per-group model access.
Global tool policy
Deployment
Open to members — everyone's agent can call it. Admins only — members are blocked. Disabled — nobody, admins and shared agents included. This is the ceiling — groups can only tighten it further. Everything is open by default on this trusted deployment.
Usage
Deployment
Activity across the platform — counts only, no conversation content.
Recent model work
Estimated Claude API chat cost in USD, not a full bill. Auxiliary model calls, discounts and missing usage are excluded. Older events are grouped by session rather than task.
Chats by user
Models
Top tools
Channels
Audit trail
Recent errors
Logs
Deployment
Live server log (last 5000 lines kept in memory; stderr still goes to docker logs).
Apps
Deployment
Toggle an app OFF to hide it for everyone (left rail, palette, settings). Use it to remove features that don't apply to your deployment — e.g. Email when corporate mailboxes can't connect.
Models
Choose which chat models users can pick. None selected = all models available. Group-level grants (Groups pane) can tighten this further. Global admins always see every model.
Telephony
Deployment
Connected to PrismVox. Recognised callers use their Prism agent; unknown callers use the Vox switchboard. Outbound calls use the Vox outbound agent. Saved agent settings apply to new calls.
App access
Members of the selected groups can consult the reception archive and use the Telephony app. Personal calls remain private.
Agents & models
Conversation model — recognised callers
Choose a model configured in AI provider. The default follows the deployment's AI model. Applies to new calls; the caller keeps their Prism profile, memory and permitted tools.
Switchboard — unknown callers
Instructions for callers whose number does not match a Prism profile. Personal Prism memory and files are not shared with the switchboard.
Selection is saved immediately. Documents in the selected collection are available to unknown callers. Manage group documents in Knowledge Base.
Upload documents to a switchboard collection:
PrismVox model — unknown callers & outbound calls
Outbound instructions
PrismVox runs outbound missions and conversations with transfer recipients, including when connected to Prism. The mission is added to these instructions for each outbound call.
Voice & greeting
Voice & inbound greetings
The voice applies to all calls. Inbound greetings are spoken on pickup, before the first generated response.
The recognised-caller greeting is used when the incoming number matches an approved Prism profile.
Speech providers
Spoken phrases
Messages used at specific stages of a call or as fallbacks when no response is generated. Keep %s placeholders for names and the reason for the call.
Pronunciation
How the agent reads acronyms aloud. Comma-separated.
Transfers & directories
Transfer directory
Everyone who can receive a transferred call: approved accounts with a phone number on their profile. Blind puts the caller straight through. Attended asks their name and reason first, announces them, and lets the recipient decline. Each person sets this on their own profile; you can override it here.
Loading…
Outbound directory
People and businesses the agent can call — the plumber, a supplier, a customer. Ask it in plain words ("appelle le plombier et prends rendez-vous") and it looks the number up here. These contacts receive outbound calls; they are not transfer destinations.
Loading…
Call handling
Limit the number of exchanges with a transfer recipient while the caller waits on hold.
Phone connection
SIP trunk
Loading status…
"Save & apply" re-registers the trunk without restarting Asterisk.
Shared agent
Group
The agent your members mention in the room. It runs with the rights you set here.
Shown next to the shared agent in the room.
Each model call counts as one iteration; one response can request several tools. 10–500, blank = default. Raise it if the shared agent stops with "iteration limit reached" on long tasks.
Extended reasoning
Thinking mode for models that have one (Qwen3, DeepSeek-R1, gpt-oss…). Off = faster, cheaper replies. No effect on Claude models.Guided: detailed. Standard: compact (formerly Lean). Minimal: essential contracts and documentation on demand. Permissions stay the same.How much a thinking model reasons (when extended reasoning is on). Accepted values depend on the model — gpt-oss: low/medium/high, Qwen3.8-Flash-Next: low/medium/xhigh; an unsupported one is refused by the server, pick another.
On the phone — members of this group
When someone in this group phones in and is recognised by their number, this is who answers. It replaces their own agent personality for the call — their memory, profile and knowledge are untouched, and that is where the continuity lives. A phone agent has to call its tools to transfer, take a message or hang up, and a dense personality is measurably bad at that, which is why this is one text for the whole group rather than everyone's own. Blank = the built-in text.
Strangers never see this — they get the switchboard, configured in Admin → Telephony.
Webex integration
Connect a Webex bot so members can talk to this shared agent in Webex spaces — it answers when @mentioned (group spaces) or on every message (1:1). Create a bot at developer.webex.com and paste its access token.
Per-sender permissions — comma/space/newline-separated emails; * = everyone in the space.
Where scheduled jobs post (cron with deliver:"webex"). Only group spaces the bot was added to are listed — add the bot to a space, then refresh.
Microsoft Teams integration
Connect this group's shared agent to Teams. It answers direct messages and @mentions in group chats and channels. Register a bot with a single-tenant Microsoft Entra application, then enable its Microsoft Teams channel.
Set this endpoint in the bot's Microsoft configuration. Microsoft must be able to reach it over public HTTPS.
Per-sender permissions: Microsoft Entra user object IDs, separated by commas or spaces. Use * for everyone who can talk to this bot in the configured tenant. Blank denies access. Group tool restrictions still apply.
Scheduled answers with deliver:"teams" go here. Mention the bot in a group chat or channel thread, then refresh. Private chats are excluded.
After saving, download the app package and upload it in Teams → Apps → Manage your apps → Upload an app. Your Microsoft administrator may need to allow custom apps. Configuration saved here does not verify Microsoft credentials or endpoint reachability.
Knowledge base (RAG)
Group
Collections shared with the whole group: the shared agent and every member's agent search them. Members see them read-only in their settings — you curate them here. Uploading to a new name creates the collection; uploading a file with the same name as an existing document replaces it.
MCP servers
Group
Servers connected here are shared with the whole group: the shared agent and every member's personal agent can call their tools. Members see this list read-only in their settings.
Transport (Streamable HTTP / legacy SSE) is auto-detected from the URL. To authenticate, pick a stored secret — it is sent as a Bearer token, or create one inline from the picker below.
Group secrets
Group
Secrets stored here are scoped to this group only — used for its MCP servers / group tools, isolated from every other group. Members see the list read-only; only group admins add or remove them.
Group tool access
Group
Toggle a tool ON to let your group's members use it, OFF to restrict it to admins. You can only tighten the global policy, never loosen it — tools locked globally stay admin-only.