ShareAI current scopes and public assistant account

Updated 4 November 2026 against ShareAI's published scope reference.

• Network-cost inference accepts inference; Gateway inference accepts monetization_gateway. Existing surcharge grants remain supported. Identity-only and payment-only grants cannot run AI.
• New activation requests the configured scopes; sign-in continues to request identity only. An account must connect before it can sign in. Refresh never adds permissions. The legacy/current Gateway scope names are treated as equivalent when validating a refresh.
• Refresh and revocation use the grant's recorded public client ID, preserving existing public/PKCE grants when an installation rotates to a replacement client. A new client secret is never sent to a different client's token exchange.
• The model catalogue separates verified embedding IDs from chat. Availability still comes from the exact connected account's response. Do not replace the embedding model on an existing index without rebuilding its vectors; matching vector dimensions alone is insufficient.

Public brand assistant

Public visitors have no signed-in user. A tenant administrator can explicitly opt their own ShareAI connection into the public assistant by setting metadata.aiohm.public_assistant to true and selecting metadata.aiohm.model. This is an operator opt-in; it is not implied by signing in or connecting an account. There must be exactly one opted-in tenant-administrator account. Never copy a grant between tenants.

The public agent reuses CustomerFundedShareAiService, its per-account request lock, local usage thresholds, token renewal and usage events. It sends only the verified public identity, approved public RAG results, and that visitor's conversation. Requests use a 12-second timeout, 800 output-token maximum and no automatic cross-provider retry. Disabled/revoked opted-in accounts fail closed.

Failover repair

MuseMirrorAgent must delegate provider/model/temperature selection to BaseLlmAgentWithFallback. Reloading the primary provider independently ignores the execution override and retries the failed primary under a different fallback label. Ollama discovery requires a chat model, rejects embedding-only model lists, and has a two-second total budget with short positive/negative caches. RAG's separate embeddings detection is unchanged.

Verification

Focused tests cover current/legacy scopes, rejected identity-only grants, token refresh across client rotation, model classification, public-only content, bounded requests, and the actual execution override. Live OAuth tests belong only on registered hosts. Do not register development callbacks without the operator's approval.

Reference: https://shareai.now/docs/api/builders/auth/scopes/