Agents

Erebros distinguishes two kinds of provider, because they fail and behave differently.

Chat models

A chat model returns text. You give it a prompt, it gives you an answer. It has no tools and no working directory. Kimi, the local Ollama models, the Model Studio plans (Qwen, DeepSeek) and the PROVIDERS aggregators are chat models.

Agentic CLIs

An agent has tools and a working directory. It reads and writes files, runs commands, and takes several steps before it answers. Claude, Codex, Grok, Antigravity, Qwen Code, OpenClaw, OpenCode and Hermes are agentic. They drive a real binary on this machine.

One Chat, many engines

Since v6.14.0 every model engine — chat API or agentic CLI — answers in the same Chat window, selected from the engine row or by /chat?engine=<id>. The settings column redraws per engine, so Claude keeps its working directory and permissions while Kimi is not offered a directory it has no concept of. Hermes, OpenClaw and OpenCode are agents from agents.config.json rather than models: they keep their own consoles under Agents in the sidebar, and OpenClaw and OpenCode are also reachable as Chat engines.

One fleet, one switcher

Both kinds live in the same fleet, and any of them can be the global model — the model that default-engine apps (Loop, Pipeline) use when they run a one-shot prompt. A saved choice is treated as an instruction: it is honoured as-is, not silently substituted when a probe is slow or wrong. If a pick has genuinely stopped working, the failover chain answers from the next candidate and says so.

Availability

A provider is “configured” when this install holds its credentials — not when a catalogue happens to answer. Some aggregators serve a model list without a key at all, so a list is not evidence of a working provider. The distinction matters for the sidebar dots and the Model Manager.

Next

• Orchestration
• Chat — the window every engine answers in.