Add AI connection

Configure a provider once, then manage it here without touching code.

Show only models that fit in GPU

For this connection’s Ollama models, hide models without a reported file size or above the GPU limit in the chat and default-model menus. Other providers remain visible. Automatic routing is unaffected.

Enforce ZDR

For OpenRouter, require a zero-data-retention upstream endpoint.

Allow tool calling

Let the model return function calls for tools supplied by a client such as Synology. MCP access is controlled separately.

Connection enabled

Make this provider available to the model router.

Linked conversations

Select earlier chats to make their transcripts available as context. Use the short code or link to identify a chat quickly.

Chat access

Select which resources this conversation can use. Only your enabled resources appear.

Model connectionsLeave all unchecked to keep normal Hub model routing. Check one or more to limit this chat to those connections.
Email and calendar accountsSelect your Google account here to let a private, tool-capable model read Gmail and upcoming calendar events. Preparing an email never sends it; you must review and approve each draft.
MCP serversOnly discovered, approved, explicitly read-only tools can run in chat, and only if the selected AI connection supports tools. Changing tools requires an agent with separate approval settings.
Internet researchSearch terms leave this server. Use a tool-capable model. Private Memory, Vault and integrations are separate permissions.

Add agent

Assignments are explicit; new connections and spaces are not granted automatically.

Add email account

Credentials and OAuth tokens are encrypted locally on Unraid.

Enable calendar access

After OAuth authorization, calendars can be discovered and configured individually.

◆
Private connection

Private content may be routed only to private models such as your local Ollama connection.

Account enabled

Make this account available to approved hub tools.

Next: authorize from the account card

Recent Gmail

Messages from the connected account.

Add calendar event

Compose Gmail

Review every message before it is sent.

Add MCP server

Use a private LAN/container endpoint. The Hub blocks public destinations and redirects.

Hub-approved tools

Save first, then use Discover tools. Select only tools the Hub may assign to agents.

No tools discovered yet.
Server enabled

Disabled servers remain configured but are hidden from agent tool lists.

Delete connection?

New client API key

Copy this now. Rotating the key immediately invalidates the previous one.

Memory-space contents

Select the associated items you want to remove. Nothing is deleted until you confirm.

Conversations
Memories

One-time OAuth provider setup

Configure each provider once. Everyone can then connect accounts through a normal login popup.

FOLLOW THESE STEPSGoogle setup
Open provider console ↗
Open beginner guide with illustrated steps ↗
    Register this exact redirect URI

    New Knowledge Vault item

    Use [[Note title]] to create graph links and backlinks.

    New memo

    The memos list shows only the first two lines unless Private is enabled.

    Unlock protected note

    Unlocking lasts five minutes in this browser session.

    Save memo to Memory

    This creates a reusable general memory. The original memo is retained.

    Add model alias

    Aliases can select automatically or pin a specific connection and model.

    Models under this hub

    Select the models this hub can use. Sharing the hub automatically shares these models. An empty selection stops routing; it never falls back to other models. Existing automatic hubs must be given a model selection before sharing. These model selections control this hub for everyone; saving the pool does not grant a user access until you click Save hub access.

    ◆
    Private only

    This alias may never route to a cloud-standard connection.

    Alias enabled

    Change password

    Use at least 12 characters. Initial passwords must be changed before normal use.

    Share conversation

    Create a view-only link. Shared views never import into anyone's memory.

    Connect this project to Gitea

    Keep building locally or connect a private repository to sync your work.

    Connecting imports missing files. Existing files with different content require resolution before connecting. Sync publishes your workspace to the selected branch.

    Add a project

    Imported files are copied into a server workspace. Dependencies, generated output, and common secret files are excluded. Use .aihubignore for additional exclusions.

    Your BuildX connections

    These connections belong only to your signed-in Hub account. Other Hub users use their own connections. Saved tokens are encrypted and are not displayed again. Credentials are never available to the coding model or task commands. Use service-account API tokens here, rather than account passwords. Mini environment ID is 2. Save connections before testing.

    Set up dedicated BuildX accounts

    Gitea: create a regular service user, grant Write access to each project repository, then create a repository-write/user-write access token in that user’s Settings → Applications. For automatic personal repository creation, use the service username as the default owner. Existing repositories retain their own owner; organizations require organization-write scope and creation rights.

    Portainer: create a regular service user and a buildx-deployers team, grant access to Mini, and share project stacks with that team. Sign in as that user and create a token under My account → Access tokens. Use server URLs without /api at the end; Mini’s environment ID is 2.

    Configure project commands and the deployment Compose file in Project settings. Review checks, then use Publish to accept changes, create a version and deploy. A successful connection test alone does not verify a build or deployment.

    
    
    Release reviewed work
    After creating the release, use Deploy an environment in Publish for reviewed deployment and route creation.
    Only accepted changes can be released. A successful stack request does not prove application health.
    
    
    
    
    
    
    
    
    
  
  
    
    
    
    
    
  
    Confirm BuildX action
    
  
    New project file
    Paths are relative to this project. Save the file after entering its contents.
    
  
    Model reply tester
    Models run one at a time, including models marked unavailable. Embedding and reranking models are listed separately. A failed check may be temporary; review or retry before disabling a model.
    History retains dated runs. GPU / CPU percentages describe local Ollama runtime memory placement, not processor utilization. Cloud: not applicable; other providers: unavailable. Response time excludes waiting in the Hub queue.
    Queued message
    Memory
    Share memory space
    Choose access for each person. Read only lets them view and use general memories. Can add memories lets them contribute. Only you can manage sharing, delete memories or change cloud permissions.
    Private chat history, Memos and Vault content are not shared. Removing access stops new requests; it cannot remove information someone already read.
    Master requirements
    Replace this conversation’s complete brief. Previous versions and files are preserved. Saving does not start the AI.
    Edit queued message
    Only waiting messages can be changed. Already applied instructions remain in the conversation.