Skip to content

Model access and usage

Use Settings → All projects → Models to connect providers, choose default models and reasoning levels, and control which models appear in pickers. Agents is a separate Settings page for decision models, automatic model choice, compaction, and Code Mode.

  1. In Models → Available to connect, find your provider.
  2. Select Sign in and follow the provider’s prompts. Providers may use a browser, a device code, or an API key.
  3. In Default models → Board chats, choose an available model.
  4. Optional: Set a different Ticket Sessions model for ticket work.

The provider appears under Signed in after you connect. If a request later fails to authenticate, repair the connection there, then recover the Session.

New chats use the composer’s explicit model choice, then the project’s override, then the matching app-wide default. Changing a default does not change an existing Session’s selected model. You can change that Session’s model through its own model picker.

For a project override, select the project in Settings, then Sessions → Chat → Model. Configure → Sessions opens the same page. Choose the model and its reasoning level. This applies to new Board and Ticket chats in that project. An Automation’s Project default uses the same override.

Models → Default models has six app-wide rows. Each stores a model and its reasoning level. A named tier uses its app-wide row rather than the project’s model override.

Row Used for When unset
Board chats Board Sessions Choose a model here; this is the root default.
Ticket Sessions Ticket Sessions Uses Board chats.
Fast Quick, low-cost work Uses Ticket Sessions.
Deep Complex reasoning, planning, and review Uses Ticket Sessions.
Visual Work with images, screenshots, and pages Uses Ticket Sessions only if that model accepts images. Otherwise, choose a model that does.
Utility Chat names and summaries Uses the chat’s own model.

Choose a model, then select one of its available reasoning levels in the same control. Visual offers only models that accept images.

Signed in lists your connected providers. Available to connect lists other providers. Use Catalog to hide or show individual models in pickers. Hiding a model does not remove its provider connection. Refresh models checks for catalog changes.

A decision model answers typed questions about structured data. It can pick an option, score something on a scale, or answer yes or no, with probabilities and confidence. It does not write prose.

Agents can use the classify tool for repeated decisions, such as labeling a list or checking a page’s state. Configure it in Settings → All projects → Agents → Decision model:

  • None is the default. Decision-model calls are disabled.
  • Local uses a llama.cpp llama-server on this Mac. Enter Server, such as http://127.0.0.1:8080, and optionally Model id for a multi-model server. Select Test, then Save. Only loopback addresses are accepted.
  • Cloud uses a provider-hosted classifier. Choose Cloud model, review the disclosure, and select Allow. If it shows Needs setup, use the provider’s sign-in action to connect it on the Models page.

The decision model does not approve tool calls or change a Session’s authority.

A local classifier receives its input on this Mac. A provider-hosted classifier receives the data being classified. For classify, that is whatever the agent passes to it, which can include Session text, tool results, and page content. Provider credentials stay outside prompts, transcripts, and tool arguments.

Choose None to stop calls immediately. A Session that already has classify keeps the tool, but calls report that no decision model is configured. The agent then decides for itself.

A project can choose another classifier or None in its Sessions → Chat → Decision model row. To disable classifiers everywhere, also remove or disable project overrides. Use the row’s reset control and Revert to global default to inherit the app-wide choice again.

Sessions created with a decision model receive classify. Starting without one does not give the Session that tool. Decision calls count toward Session usage and volli cost.

With a Cloud decision model, Pick models automatically is off by default. Enabling it asks permission to send the first message of a new chat, delegated task, or Automation Run, along with the configured model choices.

Automatic choice applies only when no model, tier, or reasoning level was specified. It chooses from your configured defaults and tiers, using their saved reasoning levels. It never overrides an explicit choice.

The choice happens when the Session starts. An unavailable, slow, or uncertain classifier falls back to the configured default. Auto-picked in the Session’s model picker shows the selection and alternatives. You can still choose another model yourself.

A project with its own provider-hosted decision model has its own Pick models automatically switch under Sessions.

In Settings → All projects → Agents → Compaction, Automatic compaction is on by default. It summarizes older model context as a Session approaches its context limit. The Session’s durable history remains available.

Turning it off does not disable /compact or recovery after a provider rejects a turn because the context is too large.

Settings → All projects → Agents also contains Code Mode’s Default and Exceptions controls. Code Mode lets a Session run a short JavaScript program that combines calls to its own tools. See Code Mode for the available modes and limits.

Volli shows two kinds of usage:

  • Session usage totals recorded model operations. A cost is marked Provider-reported, Estimated from the catalog, or Unverified. A partial total means some operations could not be priced. Estimates are not invoices.
  • Usage limits, beside ⌘K, shows subscription allowance windows reported by supported providers. These are separate from Session cost totals.

Settings → All projects → General → Show cost and token usage controls whether usage appears in the interface. Use volli cost to inspect recorded usage from the CLI.