Threads & approvals
Conversations with agents where every step is visible and risky actions wait for you.
A thread is an agent as a conversation instead of a terminal. You see answers, every tool step (reading or changing a file, running a command) and the agent's questions as cards.
Supported engines
| Engine | Connection | Model choice |
|---|---|---|
| Claude Code | stream-json | model and thinking effort |
| Kimi Code | ACP | model and thinking effort |
| Hermes | ACP | every model of your provider (e.g. via OpenRouter) |
Approvals
By default threads only ask for risky actions: guardrail rules (push, delete, deploy …), files outside the project folder, and questions from the agent. Editing project files and harmless commands run without asking. Then an approval card appears: Allow, Always allow or Deny (with a reason). Multiple-choice questions are answered right in the card.
Prefer confirming every action? Turn off Settings → Guardrails → “Threads: only ask for risky actions”.
On top of that, guardrails check input against your rules.
Working in a thread
- Attachments: paste screenshots with Ctrl+V or drag files in. Images show as previews (click = full size); agents that understand images get them directly.
- Copy and forward: under every message you can copy the text or send it to another agent with Send to agent. Long text (e.g. pasted code) arrives there as a file instead of a wall of text.
- Prompt library: Ctrl+R opens saved and earlier prompts.
- Usage: tokens, cost equivalent and context usage per thread.
- Handoff: hand a thread to another engine — the new agent gets a summary of the conversation.
- Custom instructions apply to all threads, see Settings.