NtwineAgents
Ntwine runs Claude Code, Codex and ACP agents on your own sign-in.
Ntwine doesn't bring a model of its own. Each thread is driven by an agent CLI you already use, under the account you're signed in to, so it's your subscription, your rate limits and your settings. You pick the agent per thread, and you can run several against the same repo at once.
Claude Code and Codex
Ntwine runs its own pinned copy of each CLI, the exact build it was tested against, rather than whichever version is on your PATH. It downloads them once, on first start. To point it at a different build, set NTWINE_CLAUDE_BIN or NTWINE_CODEX_BIN.
If you're not signed in, Ntwine shows you the command to run in your own terminal (claude auth login or codex login) and picks up the credential once you have.
ACP agents
Any agent that speaks the Agent Client Protocol can run as a thread too, from your PATH:
| Agent | Command |
|---|---|
| OpenCode | opencode acp |
| Cursor agent | cursor-agent acp |
| Gemini CLI | gemini --acp |
| GitHub Copilot CLI | copilot --acp |
| Kilo | kilo acp |
| Qwen Code | qwen --acp |
| goose | goose acp |
| Kimi CLI | kimi acp |
| Grok Build | grok agent stdio |
| pi | pi-acp |
An ACP agent's models show up in the model picker beside it, grouped by provider.
Model and effort
The picker beside the prompt box sets the model and effort for the next thread, and Make default keeps your choice. Effort defaults to auto, where a quick model reads each prompt and picks the level it needs, from low up to max. You can change a running thread's model from its own prompt box.
Ntwine's own model calls
Ntwine doesn't have a model of its own, but it does ask one for four small jobs:
- naming each new thread
- writing a one-line status of what each thread is doing
- choosing an effort for auto
- reading a schedule out of a prompt
Each one is a short call on your own Claude sign-in. Settings → Background summaries turns all four off at once. Threads are then named from your prompt, show no status line and start on a fixed effort, and a prompt is never read for a schedule.
Permissions
New Claude Code threads launch in Auto permission mode. Safe actions run, and risky ones ask you to approve them in the thread. Bypass never asks, so nothing waits on you, but nothing gets checked either. You can set it under Settings, or per thread beside its model.
Usage
The quota chip shows how much of your plan is left, and warns you when it's nearly out while threads on it are still running. With Fable fallback on, new Claude threads start on Fable once the rest of your usage is nearly gone, and a thread paused by a usage limit restarts on Fable instead of waiting for the reset. A thread paused by a limit can be resumed once the limit resets, and you can resume all of them at once if you like.