Operator Models
The platform includes an AI interpretation layer that converts raw coding agent terminal output into clean, readable chat messages in your browser. This layer uses two AI model tiers and must be configured separately from the coding agent model.
Go to Settings → Operator Models.
Why this exists
Coding agent CLIs (Claude Code, DeepAgents) run as terminal applications with rich text output — animations, colour codes, interactive prompts. The platform's interpretation layer reads this output and converts it into the clean markdown messages you see in the chat interface, and detects when the agent is asking you a question.
Without this layer configured, raw terminal output is passed through directly — readable but unformatted, and agent questions will not be detected.
Two tiers
| Tier | Role | Recommended model |
|---|---|---|
| Screen | Processes terminal output in real time as the agent runs | Fast, low-cost (e.g. Claude Haiku) |
| Response | Parses the agent's complete response when it finishes a turn | Capable, accurate (e.g. Claude Sonnet) |
The Screen tier runs continuously while the agent is active, so a faster and cheaper model is appropriate. The Response tier runs only when the agent finishes a turn, so accuracy matters more than speed.
Configuration
For each tier, set:
- Provider — which AI provider to use
- Model — the specific model
- API key — leave blank to use your global key from the Model Library for that provider
Recommended setup:
| Tier | Provider | Model |
|---|---|---|
| Screen | Anthropic | Claude Haiku 4.5 |
| Response | Anthropic | Claude Sonnet 4.6 |
What happens if not configured
- Terminal output appears as plain text without markdown formatting
- The agent's questions are not surfaced as interactive prompts — you may see them as text but cannot click to answer
- The agent state indicator (thinking / executing / idle) may not update correctly
Configure at least the Response tier before using interactive mode.