Models and providers
Tau uses your configured AI providers to run agents. A provider account supplies access to models; a model tier chooses which models an agent tries, in order. Start with the bundled tiers, then change them when you need a different provider or level of reasoning.
Connect a provider
Section titled “Connect a provider”Open Settings → AI Providers in your Tau workspace. Provider accounts and model choices belong to that workspace. If the page is unavailable or read-only, ask a workspace administrator to connect the account or grant the relevant access.
Open Settings → AI Providers on your Tau instance. Credentials are stored encrypted by the server. Your installation must have its encryption key configured before you can save an account; see Runtime and access if setup reports an encryption-key error.
- Select the provider you want to use.
- Choose Add account. Enter an API key, or use the provider’s sign-in option when offered.
- For browser sign-in, follow the authorization instructions and return to Tau. Complete any code-entry step Tau displays.
- Check that the account is enabled. If you have several accounts for a provider, put them in the order you want Tau to try them.
Use Add account to connect a different account. Reauthorizing an existing account refreshes that account’s login. Provider availability and sign-in methods depend on the provider catalog installed on your instance.
Viewing provider settings requires provider-auth:read; changing provider accounts requires provider-auth:write. Connecting an AI provider does not also connect GitHub or other work integrations.
Enable the live Assistant
Section titled “Enable the live Assistant”Live Assistant sessions use OpenAI API services, a separate connection from agent model accounts. Signing in to a ChatGPT subscription under AI Providers does not configure that service.
- Open Settings → Assistant & Memory.
- Under OpenAI API services, save an OpenAI API key and enable the connection. The same connection is available under Integrations.
- Turn on Voice assistant. This enables the Realtime Assistant; you can type or use its microphone.
- Open the Assistant and try a short text exchange before enabling your microphone.
These sessions use the API account and incur API usage charges. They do not change the models used by your squad agents. If the setting or credential controls are unavailable, ask an administrator to configure them. Leave the feature off to use the text-only User Assistant path.
Choose a tier
Section titled “Choose a tier”Open Settings → Agent Types → Model tiers. Each tier contains a primary model followed by fallback models. The bundled starting choices are:
| Tier | Intended use | Primary model | Reasoning |
|---|---|---|---|
| Fast | Fast, economical work | openai-codex:gpt-5.6-luna | low |
| Standard | Balanced engineering work | openai-codex:gpt-5.6-sol | medium |
| Deep | Judgment and review | openai-codex:gpt-5.6-sol | high |
| Exhaustive | Maximum-depth analysis | openai-codex:gpt-5.6-sol | xhigh |
These are the bundled configuration defaults, not a promise that every account can use every model. Your administrator may have changed them. The tier editor shows your instance’s actual chain and any unavailable-provider warnings.
The bundled fallback order is Anthropic, then Z.ai. Fast and Standard use Claude Sonnet 5; Deep uses Claude Opus 5; Exhaustive uses Claude Fable 5. Each ends with GLM 5.3. You need a configured, enabled provider account for a candidate to be usable. A connected OpenRouter account can also supply the optional universal fallback shown in Settings.
To change a tier, choose its provider, model, and reasoning level, then arrange the candidates in the desired order. Tier edits save as you edit. Use Used by to see the affected agent types. Changing tiers requires agent-types:update.
Apply a tier to an agent type
Section titled “Apply a tier to an agent type”- Open Settings → Agent Types → Types.
- Open the type you want to change.
- Select its Tier. Leave Override chain (optional) empty to inherit that tier.
- Choose Save and start a small new task to check the result.
Tau resolves the model choice in this order: an individual agent override, an agent-type override, the type’s enabled tier, then the instance default. A type override therefore prevents later tier edits from changing that type’s model choice.
An explicit model entry uses provider:model-id:reasoning, such as openai-codex:gpt-5.6-sol:medium. A chain is a comma-separated list of those entries. The model picker is the easiest way to use the IDs available on your instance.
Use a compatible or local model
Section titled “Use a compatible or local model”In AI Providers, use the custom provider setup to enter an OpenAI-compatible API base URL, a provider ID, and the exact model ID. Supply an API key if the server requires one, then run the capability probe before assigning tiers. A model that fails the tool-calling probe cannot be assigned as Primary: Tau agents need tools to perform work.
The model endpoint must be reachable from your Tau server. A model listening only on your laptop’s localhost is not reachable from a Cloud workspace.
Use an endpoint reachable from the machine or container running Tau. Its localhost refers to that environment, which may differ from the browser or laptop you use to open Settings.
When a model will not start
Section titled “When a model will not start”Check the account’s enabled state, model ID, and tier warnings first. Tau skips candidates with missing authentication, disabled providers, unknown models, or a provider in cooldown. If none is usable, connect or correct one of the providers in the chain. If a provider is in cooldown but its limit window has already reset, clear the record with the Reset button on that account’s row in AI Providers settings, or tau provider-auth reset <provider>; Tau re-marks exhaustion on the next failure. If Tau reports that no model chain resolves, select a tier or explicit override for the agent type.
Continue with Agent types or try your choices on your first task.