Configuration and Integrations

One thing, three names. In these docs we call your assistant a Digital Twin. You will also see it called an Instance — on the Instance Settings panel, in the admin Instances list, and in your web address (your-instance.praxislxp.com) — and an Institution, which is what our software calls it behind the scenes (in API fields such as institution). All three mean exactly the same thing: one assistant, with its own knowledge, voice, members and settings. Which name you see depends on the screen you are on, not on what the thing is. An Account is different — it is the organisation that owns one or more Digital Twins.

The Instances editor is where a Digital Twin is configured (Instance is the older name for a Digital Twin that this screen still uses). It is organised as a tab bar across the top, and within each tab a set of section headings (with a quick-access rail to jump between them). This page mirrors that structure exactly, so a heading here is a heading you can see on screen.

Looking for the Instance Settings panel that twin admins open from [+] → Settings in a conversation — with the Integration Wizard, Billing, Voice, Knowledge and the other screens? It has its own page: Instance Settings. This page covers the full editor in the admin console.

Configuration and Integrations

Twin details, membership, AI behaviour, quotas, and every third-party integration.

Personalization and AI Models

Identity and branding, plus the model assigned to each job the twin does.

Knowledge and tools

RAG and Files, Connector MCP and Tools, Assistants, Onboarding Questions.

Twin Memory

Long-term Memory and Shared Memory.

How the editor is organised

This editor is available to Digital Twin Instance Admins. Depending on your entitlements, some fields are read-only, and a few are managed by the Praxis AI team.

Two tabs are always present. The remaining tabs appear only after the Digital Twin has been saved — when you are creating a new one you will see just the first two, and the rest unlock once it exists.

Tab	What lives there	Documented
Configuration and Integrations	The twin’s identity, membership and credits, AI behaviour, quotas, voice entry points, and every external service	On this page
Personalization and AI Models	Name, avatar, branding, UI, and the model chosen for each job	Personalization, AI Models
Connector MCP and Tools	External MCP connectors and the platform tool policy	MCP Servers, Tools
Onboarding Questions	The guided questions new users answer on first run	Onboarding Questions
RAG and Files	Retrieval settings, the ingestion pipeline, and the IP Vault itself	Knowledge & RAG
Assistants	Which assistants users may reach, and who may author them	Assistants
Agents	Autonomous agent workspaces. Appears only when the twin is agent-eligible	Agents
Twin Memory	Long-term Memory and Shared Memory	Twin Memory

Within a tab, use the quick-access rail beside the form to jump straight to a section heading rather than scrolling. On narrow screens the rail collapses and the headings stay in document order.

Instance Management

A Digital Twin is the thing you configure here: its knowledge, its personality, its people and its settings. The admin screens call it an Instance.

Start by clicking the Admin dashboard, located at the top of the main user interface. The Instances tab lists every Digital Twin you are entitled to manage, and lets you create a new one if you are entitled to.

Viewing Instances

List of Instances

The Instances tab lists every Digital Twin you have access to, with its name, status, and other details.

Instance Details

Clicking a Digital Twin opens the editor described below — configuration, members, and activity.

Depending on your administrative rights, your editing privileges may be limited.

Tab: Configuration and Integrations

This tab carries two top-level sections — Configuration and Integrations.

Configuration

Instance Details

Instance Name

You name a new Digital Twin when you create it. Use a consistent convention:

Canvas integration: instance_name.school_name.instructure.com
Praxis LXP integration: instance_name.praxislxp.com

To rename an existing Digital Twin, contact the Praxis AI team at humans@praxis-ai.com.

Instance Credits

The credit balance for this Digital Twin. New Digital Twins receive 50 starter credits. Purchase additional credits at https://pria.praxislxp.com/my-profile/pricing or contact your organization’s IT administrator for enterprise options.

To adjust this field, contact the Praxis AI team at humans@praxis-ai.com.

Status

Controls whether the Digital Twin is available:

Active — The Digital Twin is operational for all users
Inactive — The Digital Twin is temporarily disabled

To adjust this field, contact the Praxis AI team at humans@praxis-ai.com.

Public ID

Unique identifier automatically generated when the Digital Twin is created. Used for API integrations, Chat Completions, and system connections. It is public by design — safe to share in embed code and links; it is not a credential and grants no access on its own. This field is read-only — use the copy button to copy it to your clipboard.

Contact Emails

Specify administrator email addresses (comma-separated) for monitoring and notifications about this Digital Twin. Used for moderation alerts and system notifications.

Account

Links this Digital Twin to an Account (your organisation) for centralised credit management across several Digital Twins. Select from available accounts in the dropdown.

Access & Membership

Allow React Component Embed

Lets this Digital Twin be embedded inside another web application as a component, rather than only through a link or the standard embed. Leave it off unless a developer is building the twin into your own product; see Embed your Digital Twin.

Allow Joining

Controls how new users can join this Digital Twin:

Disabled — Users cannot self-join; they must be added by an administrator
Account only — Only users within the same Account can join
Everyone (public) — Anyone can join this Digital Twin

To adjust this field, contact the Praxis AI team at humans@praxis-ai.com.

Joining Admin Only

When enabled, only administrators can join the Digital Twin — regular users are blocked from self-joining. This option appears only when Allow Joining is set to “Account only” or “Everyone (public)”.

To adjust this field, contact the Praxis AI team at humans@praxis-ai.com.

Pool Credits

Enable credit sharing across everyone using this Digital Twin. Each interaction consumes credits based on complexity. When credits are depleted, users cannot interact until credits are replenished according to your service agreement.

When Pool Credits is disabled, requests from users draw from each user’s personal credit pool instead. To change this field, contact the Praxis AI team at humans@praxis-ai.com.

Credits Awarded to New Users

The number of credits automatically granted to new users when they first join this Digital Twin. Set to 0 to disable automatic credit awards. This field is only visible when Pool Credits is disabled (i.e., users manage their own credit balance).

To change it on an existing Digital Twin, contact the Praxis AI team at humans@praxis-ai.com.

Grant Entitlements to New Admins

Automatically grants full entitlements to new administrators, providing complete dashboard access and customization capabilities. For specific entitlement questions, contact humans@praxis-ai.com.

Prevent Switching to other Instances

When enabled, users are locked to this Digital Twin and cannot switch to any other. Useful for kiosk-style deployments or single-purpose integrations.

Disable Add Credits

Hides the option for users to purchase or add credits to their account. Use this when credit management is handled exclusively by administrators or through organization-wide agreements.

Prevent Personal Account Creation

When enabled, members of this Digital Twin are not offered a personal Pria account, and the option is unavailable to them.Two boundaries are worth knowing:

The setting applies while a member is working in this Digital Twin. Someone who also belongs to another Digital Twin without this setting can still create a personal account from there.
It covers members creating an account for themselves. Administrators adding or editing people through the admin screens are not restricted by it.

AI Behavior

Enable Moderation

Default: on. Screens every user message for unsafe content before it reaches the AI model. Flagged messages are blocked with a safe response, and an alert email goes to the Digital Twin’s Contact Emails.The model that does the screening is picked under Personalization and AI Models → System → Moderation (for example omni-moderation-latest). For the categories, what users see when a message is blocked, and how to review flagged conversations, see Content Moderation.

Cite Sources

When enabled, the Digital Twin always includes source citations in its responses. This instructs the AI to reference the documents, files, or knowledge base entries it used to generate each answer, improving transparency and traceability.

Don't prompt to personalize

Disables the initial personalization popup for new users, allowing direct access to the conversation interface without being prompted to set preferences first.

Use Location

Enables location-based responses for queries about local services, weather, campus information, and regional content. Improves response accuracy and relevance for location-dependent requests.

Compact History

Default: on. Before each request, Pria has the summary model condense the earlier part of the conversation, so long conversations keep their context while using fewer tokens. Leave it on for long, ongoing conversations (research projects, multi-session tutoring); short Q&A barely needs it.

Document Search Affinity (under RAG and Files) has nothing to do with shortening conversations: it sets how closely a passage must match a question before Pria uses it. See Knowledge & RAG Configuration.

Transparency toggles

Five independent switches decide what users see alongside each response. All default to on; turn them off for kiosk-style deployments where the conversation should feel chat-like rather than tool-driven.

Display RAG/KAG Search Details — which passages were retrieved, and their relevance scores
Display Thinking Details — the model’s reasoning trace when thinking is enabled
Display Thinking Execution — live progress while the model is reasoning
Display Tool Execution — real-time progress as tools execute
Display Tools Details — the complete agent step-by-step breakdown

An always-on autonomous agent workspace for a Digital Twin is set up by the Praxis AI team. To ask for one, contact humans@praxis-ai.com; once it is on, the Agents tab appears in this editor.

Subagents

Lets this Digital Twin answer with a team instead of a single model: one model becomes the manager, plans the work, and brings in temporary specialists — a researcher, a builder, a checker — that run in parallel and report back. Best for jobs with real steps; slower and dearer than a single answer. See Single Model vs Agent Mode for what members experience.

Enable Subagents (experimental)

The master switch. While it is off, every conversation in this Digital Twin is single-model — the default everywhere.

Turning it on makes subagents available; it does not turn them on for anyone. Each member opts in for themselves from the chat’s [+] → Subagents menu and chooses how much of a team to allow. Everyone who doesn’t opt in keeps single-model conversations.

Admins only

Limits that opt-in to administrators of the twin while you try the feature out. Members keep single-model conversations and don’t see the Subagents entry in the [+] menu at all.

Max Concurrent Subagents

Default: 10.How many helpers may work at the same time for one member. Lower it to smooth out spend and load; raise it when jobs split naturally into many independent pieces.

Subagent Tool-Call Budget

Default: 20.The most tool calls a single helper may make before it is stopped and has to report what it has. This is per helper, and separate from Tool calls per answer (cap), which bounds an ordinary single-model answer.

Subagent Resume Attempts

Default: 3.How many times a helper that fails is retried automatically before the problem is handed back to the manager. Raising it makes runs more resilient but slower to give up.

Subagent Max Tokens

Default: 0 — each role uses its own default.Caps how long a single helper’s reply may be. Set a number only when you want a hard ceiling on every role at once.

Subagent turns are billed differently. A single-model answer uses the flat rate; a turn handled by a team is billed on what the team’s work actually cost, added up and rounded once for the whole turn. A request that becomes dozens of model calls costs noticeably more. See Plans and Credits.

The models each role uses are set on the Personalization and AI Models tab — see Subagent role models.

User Quotas

Per-user credit caps for this Digital Twin, shown alongside any caps inherited from the parent account and a live count of how many users are currently at their limit. Caps limit how much a single user may consume without affecting the Digital Twin’s total balance. See Credit Management for how caps interact with pooled credits and account-wide limits.

Realtime / Convo

The entry points to live voice. Provider choice, voice selection, and avatar tuning live in the Personalization and AI Models tab — see Realtime Voice & Avatars.

Enable Convo (Speech-to-Speech)

Activates real-time voice communication. Users can speak directly to the Digital Twin and receive verbal responses in real time. See Convo Mode.When enabled, the additional Convo settings below appear.

Convo Admin Only

Restricts Convo Mode to administrators — regular users will not see the Convo button.

Convo Enable Text Input

Allows users to type text messages during a Convo session in addition to speaking.

Convo Show Transcript

Shows the live transcript of the conversation beneath the voice panel, so people can read along as they and the twin speak. Turn it off for a cleaner kiosk or demo screen.

Avatar Fullscreen

Lets people expand the avatar to fill the screen during a voice conversation.

Restrict use of assistants

Set it in Instance Settings → Voice. Off by default: every assistant of the Digital Twin can be used in Convo Mode, and a conversation that already runs an assistant keeps it when someone starts talking. Turn it on to limit Convo Mode to the assistants you check in the list below the switch; members can then switch between those assistants during a call. With the switch on and nothing checked, every assistant stays allowed.

Lock full-screen conversation

Keeps people in the full-screen avatar view for the whole conversation — the collapse button and the Escape key are turned off, and ending the call is the only way out. Use this for kiosk screens and embedded experiences.

Start Immersive Mode by Default

Launches the voice conversation automatically after sign-in — visitors land in the immersive voice experience (with a tap-to-start step) every time, instead of the classic chat.

Immersive Side Drawer

Adds a panel of your own content beside the avatar in the immersive view — a menu, a reading, a set of instructions, a form. You author what it holds, and it sits alongside the conversation rather than interrupting it. Off by default; the two settings below appear once it is on.

Drawer Side — whether the panel opens on the left or the right of the avatar.
Drawer width (px) — how wide the panel is. Widen it for content people need to read; keep it narrow when the avatar is the point.

Let the embedding page guide live conversations

Lets the page that embeds the twin send it instructions mid-conversation — for example to steer the twin as a visitor moves through a form or a lesson. The twin answers those instructions aloud. Off unless the embedding page is built to use it.

Convo Web Access Link

A short public web address that opens this Digital Twin for people with no Pria account — for example nova produces …/twins/nova. Once saved, the full link is shown with a copy button and a scannable QR code for posters, badges, and screens.Whether that page opens straight into a voice conversation is decided by Start Immersive Mode by Default; if voice isn’t available it falls back to text chat.

Anyone with the link can talk to the twin and consume credits from the pool it bills to. There is no separate publish switch — setting the address makes the twin reachable, and clearing it takes the page offline.

Full guide: Guest Twin Links.Cloning: a clone never shares this address with its original. It gets its own, derived one (nova-copy-1 for an original nova) — rename it here before you share the clone. See Cloning a Digital Twin.

Text-To-Speech

Controls the read-aloud voice used to speak responses in text conversations (distinct from Convo, which is a live voice call).

Field	Description
Use natural voices for read-aloud	Use a high-quality neural voice instead of the browser’s built-in speech synthesis
Read-aloud voice	Which voice to use when natural voices are on; only the voices the speech model can speak are listed (for example Ballad and Verse exist only on `gpt-4o-mini-tts`)

The speech model itself is picked under Personalization and AI Models → Media → Text-to-Speech, or as Speech model in Instance Settings → Voice.

Integrations

Connect your Digital Twin to external platforms and services.

Embedding & LTI

Public Authorized URLs — the URLs authorized to embed or call your Digital Twin. All inbound requests are validated against this list.

https://<your_school>.instructure.com
https://<your_school>.beta.instructure.com
https://canvas.instructure.com

For hostname matching rules (subdomains, ports, what a broad entry authorizes), see Branding & Public URLs. LTI Contexts are populated automatically when you connect an LTI placement. The system remembers which Digital Twin to launch when a user arrives through LTI or a Canvas theme integration. To disconnect, remove the placement from this list or manage it from your profile page.

Pria MCP Server

MCP (Model Context Protocol) is the open standard for connecting AI to external systems, and your Digital Twin speaks it in both directions. Outbound — the twin calling external MCP servers as extra tools — is configured on the Connector MCP and Tools tab. The settings here govern the inbound direction and the identity signing used on outbound calls.

MCP Server Enabled

Switch for the inbound MCP endpoint: when enabled, external MCP clients (Claude, agents, partner platforms) can call this Digital Twin as an MCP source — for example to search its knowledge vault — authenticating with the MCP Server Secret below.This switch does not affect outbound connectors: the MCP servers your twin calls as tools stay active regardless.

MCP Server Secret

Unique authentication token used when this Digital Twin exposes itself as an MCP endpoint. Generated automatically; can be rotated. Configure third-party MCP clients with this secret.This secret is the only accepted credential for inbound MCP. The Digital Twin’s Public ID is not a credential and will not authenticate an MCP call.

MCP Signing Secret

When your twin calls an outside tool, Pria attaches a tamper-proof note saying which of your signed-in users the request is for, and signs it with this key. Your partner checks the signature, so they can safely build tools that return one learner’s own data — the identity comes from the login, not from anything the AI wrote.Share it through a secure channel (not email). Regenerating is an immediate cutover — partners verifying signatures reject calls until they configure the new value — so coordinate before rotating. See Verified learner identity on tool calls.

Curation Enabled

Allows an external learning platform to request curated articles from this Digital Twin via POST /api/curation/search. Used by the Praxis LXP to populate its AI Curated Content library.Switching Curation Enabled on generates a Curation Secret the first time. Give that secret to the calling platform; it is sent as Authorization: Bearer <secret>. It is scoped to the Digital Twin rather than to a person, so curation keeps working when staff change roles.To stop curation, switch Curation Enabled off. Do not blank the secret — turning the switch off refuses calls cleanly, whereas an emptied secret makes the calling platform look misconfigured instead.

To add the outside MCP servers your twin calls, see MCP Servers. To let other MCP clients call your twin, see Pria as an MCP server.

Canvas

Connect your Digital Twin to Canvas LMS so it can query courses, assignments, grades, and more through the Canvas REST APIs. Without these credentials, the call_canvas and search_canvas tools are unavailable.

Canvas Client Id

The numeric identifier from your Canvas Developer Key (e.g., 10000000000217). Generated when a Canvas administrator creates an API-type Developer Key under Admin → Developer Keys.One Developer Key can be shared across several Digital Twins in the same organisation.

Canvas Client Secret

The secret key paired with the Client ID above. Click Show Key on your Canvas Developer Key to reveal and copy it. Together these enable OAuth2 so users can authorize the Digital Twin to access Canvas on their behalf.

Canvas Faculty Access Token

A personal access token generated by a registered course instructor. Optional, but required for teacher-level operations such as posting grade comments, creating announcements, or bulk grading — actions that need instructor-scoped permissions even when the end user is a student.To generate one: in Canvas, go to Account → Settings → New Access Token.

This token grants instructor-level access to the courses the faculty member is enrolled in. Store it securely and rotate it periodically.

Canvas API Scopes

A whitelist of Canvas REST API endpoints your Digital Twin may call. Each scope follows the syntax:

url:<HTTP_METHOD>|<API_ENDPOINT>

For example: url:GET|/api/v1/courses/:course_id/smartsearchScopes must be configured in two places — on the Canvas Developer Key and here. If a scope is missing from either location, the API call is rejected.

During development, disable scope enforcement on your Canvas Developer Key to discover which endpoints your twin needs. Monitor Agent Details in your dialog history to see which API calls are made, then add those scopes here before going live. Use the Upload/Download buttons to manage the list from a JSON file.

See Canvas Scopes Reference for the full list.

In its own section, LMS sign-in (Canvas and Blackboard). When enabled, people who open the Digital Twin from Canvas or Blackboard must first sign in to that LMS — Sign in with Canvas or Sign in with Blackboard — before the twin opens, rather than being prompted only when a Canvas tool is first invoked. Blackboard sign-in also needs your Blackboard site set up for it: see Require Blackboard sign-in.

Don't prompt to enable Canvas authentication

Stops Pria from suggesting to this twin’s admins that they turn on Require LMS sign-in on launch (Canvas or Blackboard). The new admin editor calls it Stop suggesting Canvas sign-in.

For a complete walkthrough — creating the Developer Key, configuring scopes, testing, and troubleshooting — see the Canvas Pria Tools guide.

Chat Completion

Exposes this Digital Twin behind an OpenAI-compatible Chat Completion endpoint, so external tools that already speak that protocol can talk to your twin.

Field	Description
Enable Chat Completion endpoint	Turns the endpoint on
Chat Completion Model	Which model serves requests arriving on the endpoint
Chat Completion Max Output Tokens	Ceiling on response length
Chat Completion Reasoning Effort	Reasoning depth, for models that support it

Requests authenticate with a Pria sign-in token (JWT) — see the Chat Completions guide.

ElevenLabs Voice Cloning

Appears while Convo uses ElevenLabs. Connects an ElevenLabs Conversational AI agent:

ElevenLabs Agent ID (for Digital Twin Voice) — the Agent ID from your ElevenLabs agent
ElevenLabs API Key — must match the key in your ElevenLabs agent’s Custom LLM settings; without one, the platform-wide key is used if configured
ElevenLabs Connection Method — WebSocket (the default) or WebRTC

ElevenLabs also needs the Chat Completion endpoint above. See the ElevenLabs Voice Agent guide.

Cartesia Voice

Appears while Convo uses Cartesia: voice selection and tuning. See Realtime Voice & Avatars.

The Anam Avatar and Anam Meeting Bot sections of this editor are managed by the Praxis AI team. Tune your twin’s Anam avatar in Instance Settings → Voice (see Realtime Voice & Avatars); to send your twin into meetings, contact humans@praxis-ai.com and see Meeting Agent.

Anam and Cartesia voice conversations listen through Deepgram Flux. Its language is set by Speech Recognition Language (Deepgram Flux) under Personalization and AI Models → Convo (Live Voice & Avatar Calls), and applies wherever the twin listens through Deepgram — avatar conversations, embedded twins and meeting bots. English only is the platform default and the most accurate for English; pick the multilingual option for people who speak another language or switch languages mid-conversation. OpenAI GPT-Realtime, Gemini Live, ElevenLabs and xAI transcribe inside their own connection and ignore this field.

Google and Microsoft

Both sections work the same way:

Field	What it does
Enable Access	Turns the provider’s services on for this Digital Twin
Use Digital Twin identity	Everyone reaches the provider through one account the twin connects, instead of each person connecting their own. The service list and the connect button appear only with this on.
Microsoft tenant (optional)	Microsoft only: restrict sign-in to one Microsoft organization

A connected Digital Twin account takes precedence over any personal account a user connected: if the twin connects a shared Gmail, everyone uses that Gmail.

Services, permissions and the connect, reconnect and disconnect steps are on Google Workspace and Microsoft 365.

Kaltura

Connect the Kaltura Video Platform so the twin can search and reference videos hosted in your Kaltura account:

Field	Purpose
Kaltura Partner Id	Your Kaltura account (partner) identifier — in the Kaltura Management Console under Settings → Integration Settings
Kaltura Secret	The API secret shown next to the Partner Id: an Admin Secret (full account access) or a User Secret (limited, read-only)
Is Admin Secret	Turn it on when the secret is an admin secret. When it is off, Kaltura Client Id (the Kaltura user ID the secret belongs to) is also required

The Kaltura secret is a sensitive credential. It is stored encrypted and never sent back to the browser.

Amazon Web Services

Connect your own AWS services — primarily Amazon Bedrock for model access and Amazon S3 for storage — so the twin uses your AWS account under your organization’s billing and security policies.

AWS Region

The region where your services are deployed. Choose the one closest to your users or where your Bedrock models and S3 buckets are provisioned.

us-east-1 (N. Virginia) — Broadest Bedrock model availability
us-west-2 (Oregon) — Good Bedrock coverage
eu-central-1 (Frankfurt) — European data residency
ap-southeast-1 (Singapore) — Asia-Pacific

Not all Bedrock foundation models are available in every region. Check AWS Bedrock model availability.

AWS Access Key Id

The access key ID from your AWS IAM credentials. Generate this in the AWS IAM Console by creating an IAM user or role with the necessary Bedrock and/or S3 permissions.Leave both of the Digital Twin’s keys empty to use the pair set on its account (Accounts → Shared Integrations). A complete pair set here on the twin wins; a single key is ignored and the account pair is used.

AWS Secret Access Key

The secret access key paired with the Access Key ID. Shown only once when you create the IAM credentials — store it securely.

This is a sensitive credential. It is stored encrypted and excluded from API responses by default. Never share it in client-side code or logs.

IAM Permissions: Your IAM user or role needs:

Amazon Bedrock — bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream, and bedrock:ListFoundationModels
Amazon S3 — s3:GetObject, s3:PutObject, and s3:ListBucket on the relevant buckets

Follow the principle of least privilege.

API Keys

By default the Digital Twin reaches every AI provider through Praxis-managed access. When you provide your own API key, the twin sends that provider’s requests with your key — the provider bills your account for the raw usage, and you control rate limits and model availability. Conversations are still metered in Pria credits; BYOK customers buy those credits at a discount through our sales team. This is the bring-your-own-key (BYOK) section.

How BYOK routing works: when a model is selected for a Model Use and an API key exists for that model’s provider, the request goes out with your key. If no key is set, it uses Praxis-managed access. The choice is per provider, so you can use your own OpenAI key while Anthropic requests go through Praxis.

OpenAI API Key

Connects to OpenAI models including GPT-5.2, GPT-5.1, GPT-5-mini, o4-mini, GPT Image 2.5, and Whisper transcription. Also enables image generation and GPT-Realtime voice.Generate your key at platform.openai.com/api-keys.

Anthropic API Key

Connects to Anthropic Claude models including Opus, Sonnet, and Haiku. Enables extended thinking and reasoning where available.Generate your key at console.anthropic.com/settings/keys.

Gemini API Key

Connects to Google Gemini models including Gemini 3.8 Flash, Gemini 3.1 Pro, Gemini 2.5 Flash, and Gemini Live for real-time voice.Generate your key at aistudio.google.com/apikey.

xAI API Key

Connects to xAI Grok models including Grok 4.7, Grok 4.1 Fast Reasoning, Grok Code Fast, and Grok Imagine (images).Generate your key at the xAI Console.

Mistral API Key

Connects to Mistral AI models via the native SDK — Mistral Large, Medium, Small, Pixtral (vision), Magistral (reasoning), Codestral (code), Voxtral (TTS/STT), and embedding/moderation models.Generate your key at console.mistral.ai/api-keys.

OpenRouter API Key

Connects to OpenRouter, a gateway to many providers: open-weight models such as DeepSeek V4 (Flash and Pro), Qwen, GLM, MiniMax and Gemma, plus Claude, Gemini and Grok models. DeepSeek models are offered only through OpenRouter — there is no separate DeepSeek key.Generate your key at openrouter.ai/settings/keys.

Cohere API Key

Connects to Cohere models including Command A. Enables Cohere’s retrieval-augmented generation and enterprise search.Generate your key at dashboard.cohere.com/api-keys.

This section also takes a Kimi API Key (Moonshot AI), a Stability AI API Key, a Google AI Platform API Key (Gemini on Google Cloud’s Vertex AI — not interchangeable with the Gemini API Key) and a RunPod API Key (see Self-hosted models on RunPod). The ElevenLabs key sits in its own section above.

Security: All API keys are stored encrypted, excluded from API responses by default, and never exposed to the browser. Only administrators with the appropriate entitlements can view or modify them.

Tab: Personalization and AI Models

Identity and appearance, plus the model assigned to each job the twin performs.

Personalization (Digital Twin)

Sub-section	What it covers
Name & Avatar	The twin’s display name, pictures and backgrounds
Content & Behavior	Digital Twin Instructions, Use About for Welcome Screen, Persona mode and guardrails
UI & Appearance	UI CSS (custom CSS for branding), UI Theme, Guest UI (Minimum screen), Disable Clipboard for Users and Disable Audio Notes for Users (hides the audio-note microphone from standard users)

UI CSS is injected as-is into the page. Test thoroughly to avoid breaking the layout.

Disable Clipboard for Users stops non-admin users from copying AI-generated content: text selection and copy shortcuts are blocked, and the copy and share-link buttons are hidden. Administrators keep full access. Useful for exams and for content that must not be extracted.

Guardrails

Under Content & Behavior, Enable guardrail instructions turns on a separate block of standing rules about what the twin must refuse, avoid or hand off — kept apart from the Digital Twin Instructions so the twin’s personality and its limits can be edited independently. With it on, a Guardrail Instructions box holds those rules. Write them as plain, concrete limits rather than a tone of voice:

Never give individual medical, legal or financial advice — say you cannot and point to a qualified professional.
Do not discuss admissions decisions or fee waivers; direct people to the admissions office.

Guardrails are standing rules, so they are sent on every turn. Keep them short and specific — a long list costs credits on every message and gives the model more to weigh before it answers. The creation interview also asks what the twin should avoid and where it should send people instead; those answers shape the same behaviour. See Onboarding Questions and Crafting Digital Twin Instructions.

Full reference: Personalization and Branding & Public URLs.

AI Models

Each Model Use — a job the twin does — is assigned a model here.

Sub-section	What it covers
Custom Models	Bring-your-own-model entries added to the catalog
Conversation & Reasoning	The conversation model, its Fallback conversation model (answers when the conversation model fails before any reply started — quota, outage, network, credential; unset = the platform default), Conversation Completion Max Tokens, Reasoning Effort, Documents Summarization and Extended Context (1M tokens)
Convo (Live Voice & Avatar Calls)	Speech-to-Speech Model, voice, Voice Reasoning Effort, Interruption Sensitivity, VAD, noise reduction, transcription language, and the Deepgram speech-recognition language
Media	Image Analysis, Image Generation, Audio Generation, Video Generation, Audio Transcription, Video Analysis and Text-to-Speech models
Subagents	The model and effort for each subagent role — see Subagent role models
System	Embeddings Generation and Moderation. The knowledge-graph models (Knowledge Generation) are chosen by the Praxis AI team.

Full reference: AI Models, Realtime Voice & Avatars, and Platform Models.

Subagent role models

When Subagents are enabled, each role in the team can run on its own model — a strong reasoning model where the thinking happens, a cheap fast one where it doesn’t. The Subagents section has a mode switch — All modes, Advanced, Expert, Mythical — and one row per role, each with a model picker and an effort picker. A choice for one mode overrides All modes; All modes overrides the platform’s per-mode picks. Each Default keeps the platform’s choice and its label says which model and effort that is; leave rows on Default until you have a reason not to. Each picker lists only the models suited to that role.

Row	The role it runs
Conductor / Conductor Fallback (All modes only)	The manager that plans the turn and decides who to bring in, and the model used when it is unavailable
Planner, Spec Writer, Plan Reviewer	Work out what should be built, and check the plan
Coder, Tester	Build the thing, and check it. Per mode you can also pin Coder for harder tasks and Coder for games and simulations.
Designer, Writer	Visual and written output
Researcher, Preflight	Gather material, and check readiness before a run starts
Judge	Scores or compares results where a run asks for it
Agentic Loop	The repeated think-act cycle a helper runs inside
Vision	Reading images; only models that can see are offered

Default effort follows the mode: Medium in Advanced, High in Expert, Extra High in Mythical. Members can still pin their own models.

Tab: Connector MCP and Tools

Connector MCP

Outbound MCP connectors — external MCP servers this Digital Twin may call as extra tools. Each connector is authored with its own configuration and credentials. See MCP Servers and Connectors.

Tools

Governs which platform tools the twin may use.

Tools of this Digital Twin

The list of platform tools, each with its own switch, a Search tools box, category filters and Show only enabled tools. A tool switched off is never offered to the AI, and a call to one is refused. Tools whose feature is off for this Digital Twin are not listed; a line names those features. To apply a ready-made set (None, Minimal, Media, Builder, All), use Instance Settings → Tools — see Tools.

Result Size (Max Chars Count)

Default: 60000 characters.Maximum characters returned by a single tool call; longer results are truncated. Increase for Digital Twins working with large documents or detailed API responses. Decrease to reduce token usage and protect against runaway tool output.

File navigation tools

Builder tools

Default: on.Lets the AI check the web pages and activities it builds before handing them over — running their checks, taking screenshots, and reviewing the result. Turn it on for a twin that builds pages or interactive activities. Subagents need this on.

Tools Guidance

Default: off.Off, every question carries short rules for using tools. On, it also carries detailed guidance on how many tool calls each kind of question deserves (and, in voice conversations, the full tool rules). Turn it on when you need more reliable tool use — research workflows especially. It adds several thousand characters to every question, so it costs credits whether or not tools are used. Personal accounts have their own Tools Guidance switch in their Tools panel.

Tool calls per answer (cap)

Default: 0 — the platform default of 20 tool calls.Most answers may call tools this many times before the Digital Twin must write the answer from what it has gathered. When the cap is reached, the twin finishes with a written answer that lists anything it could not find under “Not found”, or with an honest line saying it could not finish, never with a reply that trails off mid-search. Set a number from 1 to 50 to lower the cap for this twin; a twin can only lower the platform cap, never raise it. Leave the field empty or at 0 to use the platform default. Lower it for twins that should answer quickly from a small vault; keep the default for research-style twins that search many folders or sources before answering.

Full reference: Tools.

Tab: Onboarding Questions

The guided questions new users answer on first run, whose answers feed the twin’s understanding of each person. See Onboarding Questions.

Tab: RAG and Files

How the twin retrieves from the IP Vault, and the vault itself.

User Access

Disable File Upload for Users and Allowed Emails or Domains — who may upload.

Ingestion Pipeline

How uploaded files are processed before they become searchable: Sanitize Chunks, Chunk Keyword Enrichment, Chunk Size (Search & Knowledge Graph), Chunk Overlap (%), and File longevity (days) for this Digital Twin’s vault.

RAG

Ignore user personal vault, Max. Files In Vault, Document Search Affinity (k-NN Vector Score %) and Max. Number of relevant articles. The embedding model is under Personalization and AI Models → System → Embeddings Generation.

KAG (Knowledge Graph)

To build a knowledge graph from this Digital Twin’s files, use Enable KAG Fusion for your Digital Twin in Instance Settings → Knowledge → Ingestion. Whether each twin may choose can be decided by its account; the switch then says so.

Files in IP Vault

The vault itself, embedded directly in the tab — upload, preview, edit segments, and remove files without leaving the editor. Full reference: Knowledge & RAG Configuration.

Tab: Assistants

User Access

Whether regular users may reach assistants at all, and whether they may author their own.

Disable Assistants for Users

When enabled, regular users cannot access the Assistant Library or use custom assistants. Administrators retain full access. Useful when the twin should present a single unified persona rather than a menu of specialists.

Disable Create Assistants for Users

When enabled, people who don’t manage this Digital Twin’s assistants can use assistants but cannot create new ones while working in it, in any app. Useful for curated deployments (curriculum, customer support).

Enable Assistants by Email Domains (ex: school.edu)

A list of email domains (for example school.edu) whose members may still use assistants when assistants are otherwise switched off above. It matches on the domain, not on individual addresses — so it grants access to everyone signing in with an address at that domain. Useful for opening assistants to staff while keeping them off for everyone else. Appears once Disable Assistants for Users is on.

Available Assistants

A table of the assistants available to this Digital Twin, where you select which ones are enabled. Switching one off retires it here without deleting it. Full reference: Assistants.

Tab: Agents

The control center of this Digital Twin’s autonomous agent — its sessions, the approval queue and an activity summary.

This tab appears only once the Praxis AI team has set up an autonomous agent for the Digital Twin. To ask for one, contact humans@praxis-ai.com.

Full reference: Agents.

Tab: Twin Memory

Twin Memory is the umbrella for what the twin knows about a person across conversations. It has two parts, each with its own section.

Long-term Memory

What the twin forms on its own from conversations, with a visible receipt for every memory it saves. Whether it is available on a Digital Twin at all, and whether memories are shared across sibling twins, are Praxis AI settings; what the twin’s own administrators control here is how far it reaches once it is on, and what users see under each answer. Once long-term memory is switched on for the twin, these settings appear:

Long-term Memory: Admins Only limits it to administrators of the twin while you try it out.
Memory Mode — the memory new conversations use: Classic (key/value) or Long-term Memory (twin memories). Members can change their own in their settings.
Display Memory Details (on by default) — whether the expandable Memory row under an answer lists the memories that were Recalled to shape the reply. Turn it off for a quieter transcript; memories are still used. Anything the twin Saved, Updated or Forgotten is always shown, so a user can always see what was kept.

The same controls are also available to Digital Twin administrators from inside the product, in Instance Settings → Knowledge → Long-term Memory (the first card of that tab). Turning the feature on or off and sharing memories across the account are set by the Praxis AI team; they show as read-only.

Shared Memory

Facts the whole Digital Twin reads — set once by an administrator and available to every user of this Digital Twin. Full reference: Twin Memory and Memory Parameters & Act As.

Knowledge & RAG Configuration — Retrieval modes, chunk handling, embeddings, KAG analysis model, upload policies
Realtime Voice & Avatars — Provider catalog, voice tuning, avatar configuration
Google Workspace and Microsoft 365 — connect the twin to Google or Microsoft
MCP Servers — Inbound and outbound Model Context Protocol setup
Twin Memory — Long-term Memory and Shared Memory
Memory & Act As — Per-user memory, shared memory, impersonation
AI Models — Provider catalogs, BYOM, reasoning effort, prompt caching
Personalization — Identity, avatar, branding, model assignments
Content Moderation — Per-message safety screening
Agents — Autonomous agent workspaces, scheduled tasks, approvals
Meeting Agent — Sending the twin into a Zoom, Meet, or Teams call
Credit Management — Pooled credits, user quotas, and caps
Onboarding Questions — Guided personalization for new users

Configuration and Integrations

Personalization and AI Models

Knowledge and tools

Twin Memory

​How the editor is organised

​Instance Management

​Viewing Instances

List of Instances

Instance Details

​Tab: Configuration and Integrations

​Configuration

​Instance Details

​Access & Membership

​AI Behavior

​Subagents

​User Quotas

​Realtime / Convo

​Text-To-Speech

​Integrations

​Embedding & LTI

​Pria MCP Server

​Canvas

​Chat Completion

​ElevenLabs Voice Cloning

​Cartesia Voice

​Google and Microsoft

​Kaltura

​Amazon Web Services

​API Keys

​Tab: Personalization and AI Models

​Personalization (Digital Twin)

​Guardrails

​AI Models

​Subagent role models

​Tab: Connector MCP and Tools

​Connector MCP

​Tools

​Tab: Onboarding Questions

​Tab: RAG and Files

​User Access

​Ingestion Pipeline

​RAG

​KAG (Knowledge Graph)

​Files in IP Vault

​Tab: Assistants

​User Access

​Available Assistants

​Tab: Agents

​Tab: Twin Memory

​Long-term Memory

​Shared Memory

​Related

How the editor is organised

Instance Management

Viewing Instances

Tab: Configuration and Integrations

Configuration

Instance Details

Access & Membership

AI Behavior

Subagents

User Quotas

Realtime / Convo

Text-To-Speech

Integrations

Embedding & LTI

Pria MCP Server

Canvas

Chat Completion

ElevenLabs Voice Cloning

Cartesia Voice

Google and Microsoft

Kaltura

Amazon Web Services

API Keys

Tab: Personalization and AI Models

Personalization (Digital Twin)

Guardrails

AI Models

Subagent role models

Tab: Connector MCP and Tools

Connector MCP

Tools

Tab: Onboarding Questions

Tab: RAG and Files

User Access

Ingestion Pipeline

RAG

KAG (Knowledge Graph)

Files in IP Vault

Tab: Assistants

User Access

Available Assistants

Tab: Agents

Tab: Twin Memory

Long-term Memory

Shared Memory

Related