Skip to content

Providers and Models

AlphaBase supports multiple hosted and local providers, including OpenAI, Anthropic, Google, Groq, OpenRouter, GitHub Copilot, Ollama, Azure OpenAI, AWS Bedrock, and compatible custom endpoints. The installed runtime and current catalog determine the offered connections and models.

Connect and verify

  1. Open Models & Providers → Providers or Settings → Providers on desktop. In VS Code, use Runtime Dashboard → Providers.
  2. Choose Connect and one of the methods offered by the provider.
  3. For API access, enter your own key; for account sign-in, complete browser authorization and return to AlphaBase.
  4. Choose an available model in the composer. Select model means none is selected.
  5. Send a small request to verify execution access.

Provider credentials are reused across projects in the same host and account. Separate profiles, accounts, and environments can have separate credentials. Environment-supplied credentials must be removed from their environment source rather than disconnected only in the UI.

Replacing a saved API key does not silently reassign newly accepted requests to the new key. Start a new request after changing the connection.

Subscription sign-in

Continue with ChatGPT, when offered by a local host, authorizes supported model use through that account. Identity sign-in and permission to use the plan are separate; complete the offered plan-access consent when needed. The host can retain separate account/workspace registrations. Available models and usage depend on the authorized connection.

Continue with GitHub Copilot uses a device verification flow: open the displayed verification page, approve the code, and return to AlphaBase. Access depends on the account, plan, and organization policy. If sign-in reports missing application configuration, contact the AlphaBase distributor. The distribution needs its own configured device-flow OAuth application; users should not supply a client secret or borrow another application’s credentials.

Claude Code (preview) connects through the official installed client and its own browser login. AlphaBase requires Claude Code 2.1.268 or later. Choose a client model in the primary picker and send a small request to check access. A recent completed request can verify that saved account/model for Cost Saver. New supported subscription accounts are included automatically; saved exclusions remain until you enable the account. Turn off Cost Saver to test through the primary model picker. Login and client aliases do not establish model access, included billing or remaining allowance. This integration still needs live provider and distribution qualification.

Grok remains available through the official xAI API with your own key. A Grok consumer subscription does not currently become an AlphaBase connection. A supported subscription integration requires a documented provider arrangement.

A subscription connection is not a general API key. It does not supply audio credentials for AlphaBase voice calls or cloud dictation, nor connect AlphaBase to the ChatGPT app’s included voice sessions. Use only methods actually offered by your provider and host. AlphaBase does not promise unlimited provider usage.

Cost Saver

Cost Saver uses your connected subscription accounts for new requests. It is off by default and separate from Auto Free’s strict no-charge policy.

The composer has three independent controls: your primary model, the Cost Saver switch, and the Subscriptions chevron beside Cost Saver. The switch stays compact when enabled; selection and status details are in the picker or tooltip. With Cost Saver off, requests follow your primary selection. With it on, new requests use the subscription selection. Open Subscriptions while Cost Saver is on or off to select an account/model, Add account, or Manage accounts. These actions preserve the switch, primary model, conversation and draft.

Choose one subscription model for each provider you want to use. Rotate accounts uses each enabled provider’s selected model across its eligible accounts, respecting priority, availability and project restrictions. Selecting a model for one provider preserves the other providers’ choices. Account rotation can move between those chosen provider/model pairs; it never picks an unselected model. Providers without a saved model choice stay out of rotation. Repeated sign-ins for one provider identity do not add accounts or allowance. Single account keeps one exact saved account and model, with no substitution. Existing single-provider Automatic choices become that provider’s saved model; existing fixed selections are preserved. Rows show provider, saved account label, model and allowance, so identical model names stay distinguishable. Matching account labels include a short connection ID. Response receipts retain that ID, with the full ID in the tooltip, even after an account is removed. An unavailable exact selection never silently substitutes another subscription. If no subscription can serve the request, manage accounts or turn off Cost Saver to use the primary model; Cost Saver does not silently switch to API billing.

New supported subscription accounts are included automatically when no saved inclusion choice exists. Existing exclusions are preserved across refreshes, restarts and reconnects. To enable an excluded account, click its Include in Cost Saver button directly in the Models picker. This changes only account inclusion and retains spending settings, project restrictions, model selection and the Cost Saver switch. Open Subscriptions → Accounts to verify models or change account settings. Standard setup uses the provider account’s existing limits and AlphaBase budgets; no separate overage confirmation is required. Advanced holds priority, project restrictions and an optional Included allowance only policy. That optional restriction requires confirmation of provider overage settings; previously confirmed restrictions are retained. Manage accounts and Add account open the same Accounts tab. Models returns to Rotate accounts or Single account selection. A verified model excluded by account settings shows Not eligible with the reason. Login alone does not establish model access or included allowance; supported access conditions vary by provider and account. Expired or unverified accounts remain visible for recovery. Unknown entitlement or remaining allowance stays unknown, never unlimited or zero usage. Ready, No eligible account and Temporarily unavailable describe the saved inventory without making a model request. Every executed request receives fresh admission checks.

Cost Saver does not classify requests as lighter or demanding. New chat, research and coding requests use a capable subscription model with the normal tools, permissions and task budgets. Already-running tasks keep their accepted model, account, permissions, budget and state. Changing the picker does not take over those tasks. Ordinary requests produce one answer.

Responses remain in the same AlphaBase conversation with account and allowance details. Connecting ChatGPT preserves an existing OpenAI API key. Provider-side chat histories are not imported. Saved subscription accounts use the AlphaBase runtime; external runtimes that cannot preserve the account binding are unavailable for that choice. Claude Code client aliases do not prove access. A recent completed request verifies only that account/model, without establishing included billing or reliable autonomous coding.

Open Open context usage → Load usage to review Provider activity. It separates reported model API charges, catalog estimates and observed subscription calls by account and model. Select dates and load again to change the report. Request counts are not premium allowance units; missing provider balances remain unavailable. API equivalent and Estimated API spend avoided are conditional estimates from retained usage and prices, not bills or guaranteed savings. Missing usage, unknown charges or incomplete history can make them unavailable. Media and separately billed tools remain outside these totals.

Task limits survive interruption. Unconfirmed requests count conservatively. An older task without recoverable accounting may need a new submission; restarting or changing accounts does not renew an allowance. For an older goal requiring a new connection, choose an available model/account and Resume with selected account on the goal card, or use /goal resume --selected-account. The goal, requirements, history, recorded usage and existing limits are retained. Ordinary Resume goal keeps the saved connection.

Voice audio uses its own settings. Subscription access does not establish free audio, unlimited usage or verified remaining capacity. Live provider eligibility and limits still apply.

Choose a model

Use the composer’s model selector to search and select an available model. Check tool support for coding tasks and image support for image questions. Use the displayed capabilities and your task requirements rather than assuming a model name guarantees access or performance.

Use Models & Providers → Models or Settings → Models to change visibility. In VS Code, use Runtime Dashboard → Models. Hiding a model does not disconnect its provider.

The chat model, Sophia’s help model, the realtime call model, and the dictation provider are separate selections. Switching the chat model does not automatically switch the audio services.

If a model is missing or a request fails, inspect the exact error, connection method, account access, and model visibility. Avoid repeatedly retrying a request that is blocked by authorization or limits.

Filter Provider Inventory

Use config-level filters:

{
"enabled_providers": ["openai", "anthropic", "ollama"],
"disabled_providers": ["openrouter"]
}

enabled_providers is an allowlist. If present, all non-listed providers are hidden.

Provider Override Schema

Each provider.<id> can override provider or model behavior.

{
"provider": {
"openai": {
"options": {
"apiKey": "{env:OPENAI_API_KEY}",
"timeout": 300000,
"setCacheKey": true,
},
"models": {
"<model-id>": {
"options": {
"reasoningEffort": "high",
},
"variants": {
"fast": { "disabled": true },
},
},
},
},
},
}

Supported common option fields:

  • apiKey
  • baseURL
  • enterpriseUrl (notably for GitHub Enterprise Copilot auth)
  • setCacheKey
  • timeout (number or false)

Custom Provider Pattern

You can define non-default providers via provider records (for example OpenAI-compatible gateways), then route models through them.

Model Search (Model-First Onboarding)

/provider/model-search groups matching models by canonical family and ranks provider options by first-party/bundled/connected status. This powers “add model first, then choose provider” workflows.

Troubleshooting

If a provider is configured but unusable:

  1. check runtime/capabilities
  2. verify env key scope and endpoint URL
  3. verify provider is not excluded by enabled_providers / disabled_providers
  4. confirm timeout/baseURL overrides are valid
  5. if the model intermittently fails, check whether it has been classified degraded from recent evidence — see Models; this self-heals as soon as the model’s recent attempts improve.

See Troubleshooting.