LLM Setup - Local, Cloud or hybrid

This is more "which LLM for which role" than a single pick. Out of the box, the built-in adapters wrap local CLI runtimes that almost always talk to a cloud provider underneath — so even "local adapter" usually means local process, cloud inference. If you want true local-only inference, opencodelocal is the most flexible because it can route to a local provider (Ollama, llama.cpp, etc.) and switch per-agent. The Manifest auto-routing guide (/guides/adapters/manifest-auto-routing) also covers mixing providers from one agent.

In practice, most teams end up hybrid by role:

CEO / strategy: Claude Opus via claudelocal. The reasoning premium is worth it for the agent that decides what everyone else works on. Manager / coordinator: Claude Sonnet via claudelocal. Engineers / code-heavy: codexlocal or claudelocal (Sonnet). Codex is cheaper for pure code work. Research / memory-heavy: hermeslocal — the only adapter that persists memory across heartbeats. Polling / monitoring: small models via opencodelocal to keep cost down.

Full breakdown with cost profile, capability table, and a worked org chart in Building a Mixed-Adapter Team (/guides/adapters/mixed-adapter-teams).

What setup are you starting from — solo dev, content team, research workflow? And which provider credentials do you already have? I (or others here) can suggest a more concrete adapter map. And curious to hear what others in the community are running — drop your stack below.