LLM providers
castiel works with any language model you can reach through one of five built-in providers, configured in Help → Settings → LLM Providers. None of them are enabled by default: the chat, sub-agents, the Kali dispatcher and the looper each need a model before they run.
| Provider | What it is |
|---|---|
| llama.cpp | Run open models locally with llama.cpp's built-in server, or point castiel at a remote one. |
| Ollama | Run open models locally with Ollama, or point castiel at a remote Ollama server. |
| OpenRouter | One OpenRouter key unlocks models from many upstream providers. castiel only ever needs this single key. |
| Cline | Reach Anthropic, OpenAI, Google and more through Cline's OpenAI-compatible endpoint, with an API key or an account sign-in. ClinePass subscription models show up alongside the regular catalog. |
| Z.ai | Zhipu's GLM model family straight from z.ai — pay-as-you-go with a Standard API key, or on the GLM Coding Plan subscription. The key type decides billing. |
WARNING
Most models will flag a penetration-testing session as cyber security and refuse to work. Models with a very low refusal rate for security work perform dramatically better in castiel. The GLM 5.x family, for example, has performed shockingly well in my testings.