LLM models
LLM in the left menu manages model configurations. One config is one callable endpoint. Create as many as you need and name them by purpose.
Create a config
Section titled “Create a config”| Field | Required | Notes |
|---|---|---|
| Display name | Yes | For you, e.g. “fast” or “production” |
| Protocol | Yes | Must match the provider’s API |
| Model ID | Yes | The provider’s model identifier |
| API key | Yes | Stored and shown masked afterwards |
| API base URL | No | For self-hosted gateways/proxies; empty uses the protocol default |
| Proxy | No | HTTP proxy when required |
| Quota URL | No | Only queried when provided |
| Max output tokens (K) | No | Caps a single reply |
| Context window (M) | No | Used to estimate context usage |
| Supports image input | No | Turn on if the model accepts images |
| System notice as developer role | No | Some models require this for system notices |
Thinking
Section titled “Thinking”Fields depend on the protocol:
- Thinking mode — off / adaptive / fixed budget / explicitly disabled
- Thinking budget — token count for fixed-budget mode
- Thinking display — summarized or omitted
- Effort — supported by some protocols; controls reasoning effort
If unsure, choose adaptive and let the model decide.
Default model
Section titled “Default model”Mark one config as the default. New sessions and every agent set to “follow system default” use it.
Why keep several
Section titled “Why keep several”- By cost: a cheap model day to day, a stronger one for hard tasks.
- By capability: a vision-capable model for image sessions.
- By environment: self-hosted gateway and the official endpoint side by side.
401 / auth failure — key incomplete, expired, or the protocol does not match.
Model not found — check the model ID and account permissions.
Replies are cut off — raise max output tokens, or check that the context window is not set too small.
No quota shown — a quota URL must be configured.