Skip to main content
There are now two model surfaces in the Agents product:
  • AI Models in the main sidebar — a standalone catalog where you can browse the live provider/model inventory, star favorites, and run direct BYOK test chats without creating an agent.
  • An individual agent’s Models tab — where you decide which models that specific agent is allowed to use and which one is the default.
Current AI Models page

​
AI Models

Open AI Models from the sidebar when you want to browse what’s available platform-wide. The current page gives you:
  • A single searchable catalog across providers
  • Provider, capability, and sort filters
  • Per-model badges for context window, vision support, speed, and pricing metadata
  • Star actions for favorites
  • Chat actions for direct test chats on your own provider key
  • A Manage AI providers shortcut back to provider connections
If a provider isn’t connected yet, it still appears in the catalog but its chat actions stay disabled until you add a key.

​
Per-agent Models tab

Open an individual agent → Models when you want to control that agent’s allowed model set and default model. That view still handles the agent-specific decisions:
  • Which connected provider/models are enabled for this agent
  • Which enabled model is the default
  • Which models appear in the chat composer for per-turn switching
Only providers you’ve connected in the Secrets Vault show up there. Removing a model from an agent does not disconnect the provider globally; it just narrows what that agent can choose.

​
Switch models mid-conversation

On the Chat tab, the MODEL dropdown next to the message input lets you pick a model for the next turn. You’d typically reach for this to:
  • Use a fast/cheap model for routine questions
  • Bump up to a more capable model for a hard one
  • Compare answers from two models on the same prompt
Each response is tagged with the model that produced it, so the conversation stays readable when you switch.

​
What models look like in configs

When you reference a model in code, in manifest.json, or in a task payload, the convention is provider/model: 
anthropic/claude-sonnet-4-6
openrouter/tencent/hy3-preview
openai/gpt-5.4
The provider prefix tells the agent engine which connected provider to use. If you give it a model that isn’t enabled on this agent, the request falls back to the default. Provider availability and default enabled models can vary by engine. For example, fresh OpenClaw and Hermes agents can start with different Anthropic defaults.

​
API surfaces

The HTTP API now exposes the same split:
  • GET /v0/models — unified model-centric catalog
  • GET /v0/models/catalog — connected providers plus their live/curated catalogs
  • GET /v0/models/enabled — models pinned for the per-agent picker
  • GET /v0/models/chats — direct test-chat history
The exact catalog changes constantly as providers ship and Pinata refreshes metadata. Prefer the live AI Models page over hard-coded lists when you want to know what’s available right now.