Besides built-in API models, VibeSpace can drive locally installed CLI coding agents as execution backends: Codex CLI, Claude Code, and OpenCode.
Why it matters
Many developers already have subscriptions or sign-ins with these CLI tools. CLI backends let you:
- Reuse the existing sign-in and default model — no second API key to configure;
- Run them under VibeSpace's unified timeline, approvals, and ChangeSets instead of reading raw terminal output;
- Mix API models and CLI backends across tasks in the same workspace.
How it works
- VibeSpace detects installed CLI tools automatically (no configuration);
- Detected backends appear in the task composer marked "· default model";
- When selected, the CLI launches with your workspace as its working directory, and its session output is normalized into VibeSpace timeline events;
- The session ID is remembered so recovery can reattach to the same CLI session.
Credential safety
- VibeSpace does not read, copy, or store any CLI credentials;
- Sensitive environment variables are filtered when spawning the CLI subprocess;
- In Request approval mode, you confirm before the CLI is even started.
Choosing between API and CLI
| Scenario | Recommendation |
|---|---|
| Already subscribed to Claude / Codex / OpenCode | CLI backend — reuse it |
| Need Chinese models (DeepSeek, GLM, Kimi…) | API backend |
| Need expert groups or scheduled tasks | API backend (CLI backends don't support them) |
| Want unified ChangeSets and restore | Both work |
Known limitations
- CLI backends do not support experts or expert groups (orchestration requires protocol support only API backends provide);
- CLI backends don't offer the
schedule_createhost tool (the agent can't create scheduled tasks on its own); - Available models are whatever the CLI itself is signed into; VibeSpace doesn't override its default model.