For OpenAI-compatible endpoints that don't support the Responses API (vLLM, LiteLLM, local proxies), set the Reasoning effort to "Off" in the chat panel. "Off" only stops sending the reasoning parameter (the request then uses the standard chat endpoint) — the model may still reason internally on its own.
Leave empty to keep the existing key.
Optional — leave empty for local models.
Context
This prompt is sent as a system message at the start of every conversation.
Auto-inject
AGENTS.md / Project Instructions
Include a project instructions file (AGENTS.md) as context.
Auto Compaction
When the estimated request size passes the threshold, older messages are folded into a summary so the conversation stays within the model's context window. Images in compacted messages are dropped.
Rough estimate, not exact. When exceeded, older messages get compacted.
The most recent N user messages (and their replies) are never summarized.
Saved.
Tools
Enable or disable tools the model can use to interact with your workspace.
Approval: Ask prompts you in the chat before each call, Auto runs without asking, Deny blocks the tool entirely. Read-only tools default to Auto; everything else defaults to Ask.
MCP Servers
Configure Model Context Protocol servers. These provide additional tools and resources to the model. The approval setting on each server applies to all tools that server exposes (Auto runs without asking, Ask prompts in the chat, Deny blocks them entirely).