Using the Chat

Sending messages

The chat input accepts text, drag-and-dropped files, pasted clipboard images, and file @mentions to bring specific content into context without having to describe it.

Model selector

Only shows providers with a configured API key — Google Gemini, OpenAI, Anthropic, Cohere, or local Ollama. Switching models mid-conversation doesn’t lose history.

Permission modes

Control how automatically the agent acts before asking for your approval:

Mode	The agent…
Always ask	pauses before any destructive action (writing a file, terminal, git push)
Accept edits	applies file edits directly; terminal and git still pause
Autonomous	doesn’t pause for anything — use with a workspace you already fully trust
Plan	only plans, never executes a real action

See Orchestrator & Subagents to understand the HITL behind this.

Orchestrator thinking

The “thinking” block shows the orchestrator’s decision before acting: answer directly or delegate to coder/search, and why. This is real transparency, not just a spinner — you can see the reasoning, not just wait for the result.

Memory across sessions

A “🧠 N memories loaded” badge appears when the agent uses persistent memories from previous conversations in that response. Manage memories manually in Settings → Preferences → Memory.

RAG citations

Answers based on indexed content bring clickable [1] [2] citations — click to see the original excerpt and source.

Multi-user (Pro)

In Pro mode with multi-user web chat, threads can be shared between members of the same workspace, with RBAC controlling who sees what.