AI models

What this page is for

Exolvra agents can run on any configured provider’s model. This page sets the instance-wide defaults and controls which models are available to the router. Per-agent overrides happen on the agent’s own edit page — this page is the floor.

Default model

The Default Model field is the one every agent uses unless it overrides the choice. The dropdown shows every model from every configured provider, grouped by provider.

Smart aliases — names like auto, fast, best, opus-class, haiku-class — auto-resolve to a specific model at runtime. Pick an alias when you want portability across providers; pick a named model (claude-sonnet-4-6, gpt-4o, gemini-2-5-pro) when you want exact control.

The usual choice is a mid-tier model from whatever provider you’ve standardized on. claude-sonnet-4-6 and gpt-4o are both good defaults. Cheaper models (claude-haiku-4-5, gpt-4o-mini) run agents fine for most work at a fraction of the cost.

Temperature

Temperature controls how deterministic the agent is. Default 0.2.

• 0.0 — fully deterministic. Same prompt gives the same answer. Good for summarization, extraction, classification, code analysis.
• 0.2–0.5 — slight variation. Good for most agent work: you get consistent reasoning with enough variety to not feel robotic.
• 0.7–1.0 — creative. Good for writing, brainstorming, marketing copy. Not good for tool-using agents because the agent might hallucinate tool arguments.
• >1.0 — incoherent. Rarely useful.

Leave at 0.2 unless you have a specific reason. Temperature set on the instance is a floor — individual agents can override.

Anonymous fallback model

This field matters for the bot widget and the public Bot API. When an end user chats with a bot through the widget, the session is classified as Bot API (see Security & cloud mode). If the agent’s normal model is a CLI provider (Claude Code, Codex, Gemini CLI) and the Bot API security settings don’t allow CLI providers, the agent’s model is swapped for this fallback.

The fallback must be an API model, not a CLI alias. claude-sonnet-4-6 and gpt-4o are fine. claude-code or codex-cli are not.

Most instances never trip this fallback because their bots run on API models anyway. It’s a safety net for the edge case where an admin accidentally pointed a widget-exposed bot at a CLI provider.

Model governance

The Governance section is off by default — every known model is available to the router. When you flip Enforce model allowlist, only the models you’ve ticked can be used. Agents that reference a blocked model get an error on their next turn.

Use the allowlist when:

• You want to prevent expensive models from being used without explicit approval
• You have a compliance requirement to only use specific providers or model versions
• You’re standardizing on a subset of models for cost predictability

The page groups models by provider with Select All / Deselect All shortcuts per provider. A common pattern is to select-all for your primary provider and leave the rest off.

Ollama discovered models

Separate section for locally-installed Ollama models. Click Refresh to discover what’s actually installed on the host (queries Ollama’s /api/tags). Tick the models you want available to agents. Ollama models are free to run and don’t consume budget, but they’re usually slower and weaker than hosted frontier models.

Temperature, model, and allowlist — how they interact

The order matters. For every agent call, Exolvra:

1. Picks a model — agent-override → instance default
2. Checks the allowlist — if enforcing, blocked models get rejected here
3. Applies temperature — agent-override → instance default
4. Routes the call

An agent whose model is blocked can’t run at all. An agent whose model is allowed but expensive will still run — the allowlist is a yes/no, not a budget limit. Use the Budget page for spend control.

Common pitfalls

Locking the allowlist before wiring budget. The allowlist stops unknown models. It doesn’t stop spend. If you’re worried about cost, enforce budget limits first; the allowlist is for compliance and governance, not cost control.

Setting temperature high on tool-using agents. Tool-calling agents need to emit structured outputs (tool names, argument schemas). High temperature breaks that. Keep tool-using specialists at 0.2 or lower.

Forgetting the fallback model is a CLI alias. If the widget is configured but failing with model errors, check that the anonymous fallback is an API model, not a CLI alias.

Over-tightening the allowlist. Agents need smart alias support to use the router’s cost-awareness features. If you tick only three specific model IDs, the router can’t downgrade automatically on budget pressure. Include at least a cheap and an expensive tier.

Where to go next

• Providers — configure the provider API keys
• Budget — limit spend, separate from the allowlist
• Agents dashboard — override the default on specific agents