Models, providers & cost

Hermes works with any OpenAI-compatible API. The docs list OpenRouter, Nous Portal, OpenAI, Anthropic, Google Gemini, z.ai/GLM, Kimi/Moonshot, MiniMax and others, plus local models through Ollama, vLLM, llama.cpp, SGLang and similar servers.

Run hermes model to add providers and pick a model. Inside a chat, /model switches between providers you've already set up.

Nous Portal is Nous Research's paid subscription service. One OAuth login gives you access to 300+ models plus the Tool Gateway, which handles web search, image generation, text-to-speech and a cloud browser without separate accounts. The docs recommend it for newcomers.

It's optional. Hermes works with your own provider keys, and you can mix the two per tool. Set it up with hermes setup --portal and check it with hermes portal info.

Yes. The easiest route is Local Models: in Hermes Desktop open Settings > Providers > Local Models, or in a terminal run hermes model and choose Local models. Hermes installs a pinned llama.cpp runtime, then prices every catalog model against your machine before you download anything: green means it fits your GPU (or Apple unified memory), amber means it spills into system RAM and runs slower, red means it's too big. A star marks the recommendation.

Hermes picks the best quantization that fits, never goes below 4-bit, and gives every recommended model at least a 64K context window, which is also the minimum Hermes needs. Nothing leaves your computer once the model is downloaded.

You can still point Hermes at your own server instead: run hermes model, choose Custom endpoint, and enter an address such as http://localhost:11434/v1 for Ollama (vLLM, llama.cpp server, SGLang and LocalAI work too). Make sure the context length you set matches what the server actually uses. Very small models can struggle with tool calling, for example claiming to save memory without doing it.

Hermes itself costs nothing. Your costs come from model usage with your chosen provider, an optional Nous Portal subscription, and wherever you host it if not on your own computer. Local models have no usage fees.

To keep bills down, run hermes prompt-size to see your fixed prompt cost, disable tools and skills you don't use, use /compress, and start new sessions instead of letting chats grow forever.

It can. The prompt cache key includes the model, so the first message after a /model switch re-reads the whole conversation at full input price. On a long session it can be cheaper to start a new session on the right model, or to delegate the task to a subagent, than to switch back and forth.

A plain /model switch lasts only for the current session unless you add --global. /model also lists only providers you have already configured; to add new ones, exit and run hermes model.

Hermes has layers of resilience. A credential pool, meaning several keys for the same provider, is tried first. Then a fallback chain switches to a different provider and model mid-session on rate limits, server overload, auth failures or connection drops, without losing the conversation.

Set it up with hermes fallback, which reuses the hermes model picker and has add, list, remove and clear subcommands. Entries are stored under fallback_providers in config.yaml, and each needs both a provider and a model.

Inside a session, /usage shows token usage, an estimated cost breakdown, context window state and, when the provider reports it, remaining credits or quota. hermes insights analyses token usage, cost estimates, tool use and activity over the last 30 days, with --days and --source filters such as telegram.

For scripts, hermes usage prints account limits for openai-codex, anthropic or openrouter without starting a session, and hermes prompt-size breaks down the fixed prompt sent on every call.

Hermes uses auxiliary models for side jobs such as context compression, session titles, image analysis, web extraction, approval checks, the skills curator and Kanban planning. By default each one runs on your main chat model, which adds meaningful cost on an expensive reasoning model.

You can route each job separately. Run hermes model and choose Configure auxiliary models, use Settings > Model > Auxiliary models in Desktop, or set keys like auxiliary.compression.provider and auxiliary.compression.model in config.yaml. The post-session memory and skill review has its own slot, auxiliary.background_review.

Subagents get their model from delegation.model and delegation.provider (or Settings > Model > Subagents), and the backup model used when your main one fails comes from hermes fallback add. A fast, cheap model is usually fine for compression, titles and approvals.

There are two documented routes. The official plugin claude-subscription-directsdk (hermes plugins install claude-subscription-directsdk) runs Hermes turns through the official Claude Code CLI on your Pro or Max plan, with no API key; it needs claude installed and logged in, and each turn counts against your subscription's Agent SDK allowance. The catalog marks it experimental.

Anthropic OAuth in hermes model routes as Claude Code against your account, but the docs say it only works on Claude Max with purchased extra-usage credits, and it uses those credits rather than your base allowance. Pro can't use that path; an ANTHROPIC_API_KEY is the alternative.

Hermes can't promise how Anthropic treats any account. Check your plan's terms and usage settings in your Anthropic account, which govern quota and overage.

First find out what's expensive. /context breaks down what fills each request (system prompt, tool schemas, skills, MCP, memory, conversation), hermes prompt-size shows the fixed cost sent on every call, and hermes insights shows usage over the last 30 days.

Then trim. Disable toolsets and MCP servers you don't use (hermes tools), start a new session with /new when you switch tasks instead of letting chats grow, and /compress long ones. Move side jobs such as compression, titles and approvals to a cheaper model under auxiliary models, and give subagents a cheaper model with delegation.model.

Avoid needless mid-chat model switches, which re-read the whole conversation at full price, and watch the cache hit rate in Desktop's status bar; a warm cache makes long sessions much cheaper. Script-only cron jobs (--no-agent) run with no model call at all.

For the main agent, pick a strong tool-calling, agentic model; Hermes runs a fast loop of tool calls, so that matters more than chat polish. The Nous Portal docs suggest examples such as anthropic/claude-sonnet-4.6 as a general-purpose agentic model, openai/gpt-5.5-pro for reasoning, google/gemini-3-pro-preview for huge context and deepseek/deepseek-v4-pro as a cost-effective coder. They also say Nous' own Hermes 4 chat models are not recommended inside the agent.

For side tasks, a fast cheap model is usually enough. The docs suggest a flash-class model for compression, titles, approvals and web extraction. The background review defaults to your main model, which reuses its warm cache; route it with auxiliary.background_review if your main model is expensive.

Small local models (roughly under 30B) often fumble tool calls, so use a stronger model at least for setup. Rankings change quickly, so test on your own tasks.

Hermes itself has no usage cap; limits come from your model provider or plan. /usage shows token use and, when the provider reports it, remaining quota, and hermes usage prints account limits for some providers without starting a session.

On transient errors (rate limits, dropped connections, 5xx) Hermes retries each provider agent.api_max_retries times (default 3) before moving to your fallback providers, and waits out a rate limit that names its reset. After retries and fallbacks are spent on an outage, agent.auto_recovery_cycles (default 5) keeps waiting and retrying; set 0 to turn it off.

If you have a fallback configured (hermes fallback add) and want to fail over faster, set api_max_retries to 0. On the free Nous tier, a session you're watching ends quickly with the time the limit lifts instead of waiting long.

Billing and account details are managed on Nous Portal itself, not in Hermes. hermes portal open opens the subscription management page, and hermes portal info shows your login, subscription and routing status.

If something is wrong with a charge or the account, contact Nous: after Send diagnostics, Desktop's error card links to Nous Portal Support, and the Nous Research Discord is the community channel the docs point to. Hermes can't change Portal billing settings for you.

If Hermes reports a billing failure mid-chat, Desktop's error card names it as billing and offers Switch provider, so you can keep working on another model while it's sorted.

In hermes model, choose ChatGPT or Codex Subscription. Hermes signs in with ChatGPT's device-code login (or hermes auth add openai-codex --browser for a browser login), stores the credentials in ~/.hermes/auth.json, uses Codex models, and can import an existing Codex CLI login.

The docs say which plan tiers qualify and how Hermes usage counts against your plan's Codex limits are not yet documented, so watch your usage in ChatGPT. Some models return an error such as “not supported when using Codex with a ChatGPT account”; that's a plan entitlement.

The Nous team has said on X that a newer “Sign in with ChatGPT” option links your plan to your Nous Portal account under Linked Accounts, with a usage cap you choose, and that the older Codex login still works. Check Nous Portal for the current flow.