Hermes Agent Models & Providers

Secrets go to ~/.hermes/.env and OAuth tokens to ~/.hermes/auth.json; the chosen model is saved in ~/.hermes/config.yaml. First-time walkthrough: set up Hermes Agent.

From the subscription table in the official providers docs:

Other OAuth sign-ins listed in the official quickstart include Qwen Portal and MiniMax. Each Bot Mode bot is its own profile, and OAuth logins are not copied between profiles — sign each one in with hermes -p <name> auth add <provider>.

What the official docs actually require and advise:

• 64K context minimum. Hosted frontier models clear this easily; local servers often default lower and must be raised.
• Tool calling matters more than chat quality. Hermes works through rapid tool-call loops — terminal, files, web, browser.
• Nous's own Hermes 4 models are not recommended inside Hermes Agent. Nous says they are tuned for chat and reasoning, not the tool-calling loop. The Portal docs point agent work at frontier agentic models from Anthropic, OpenAI, Google, and DeepSeek instead.
• Data-training warnings: since v0.21.0 every model picker warns when a selected model's tier trains on your data.
• Catalogs move weekly. For example, the v0.21.5 notes (September 24, 2026) list GPT-6 Sol/Terra/Luna and Claude Opus 5.5 newly added to the Nous and OpenRouter catalogs. Use /model to see what your provider offers today.

• Any OpenAI-compatible server (Ollama, LM Studio, vLLM, SGLang, llama.cpp): hermes model → custom endpoint, then enter the base URL (for Ollama, http://localhost:11434/v1) and the exact model name.
• Raise the context window to at least 64K — for example -c 65536 for Ollama or --ctx-size 65536 for llama.cpp, per the official quickstart.
• Managed local models: Hermes can also download and run llama.cpp for you, sizing each model to your GPU (Settings → Providers → Local Models). The docs say this Desktop interface is enabled on canary builds; other Desktop builds need the --local launch flag.
• Local inference keeps prompts on your machine, but web search, cloud browser, and messaging tools still reach the internet if you enable them.

Hardware needs depend on the model, quantization, and context length; there is no single official sizing figure. Per-runtime context and tool-calling flags, and what still uses the network: Hermes local models.

Besides the main model, Hermes runs side jobs — context compression, vision, web-page extraction, approval scoring, session titles, MCP tool routing, skill search, and more. Each task defaults to auto, which means your main model. Pointing the summarization-style jobs at a fast, cheap model is the most common override.

$hermes model
# → "Configure auxiliary models..." → pick a task, provider, model

model:
  provider: openrouter
  default: <your-main-model>
auxiliary:
  compression:
    provider: openrouter
    model: <a-fast-cheap-model>
  title_generation:
    prefer_fast_model: true   # let Hermes pick the provider's fast tier

Mid-session model switches can cost more: the next turn may re-process the conversation at uncached input rates, and Hermes warns when a switch will trigger context compression against a smaller window. Switching between Desktop and the TUI no longer rebuilds the system prompt (fixed in v0.21.2).

For what each option costs per month, see Hermes cost.