Builder mode
Models and keys
Pick the model behind Hermes, bring your own provider key, and add tool and voice keys.
SettingsModel picks the model behind Hermes and the key that pays for it. Two more sections hold your own keys for other services: SettingsTool keys for web search and browser tools, and SettingsVoice for speech.
Choosing a model
Open SettingsModel (open it). In builder mode it's under "Models" in the Settings list.
On a paid plan the card is headed "Choose what powers Hermes." Hermes comes ready to use on the built-in engine, and you can move to another provider at any time, either with your own key or by signing in with a ChatGPT or Grok account. Keys are encrypted, and you can replace them whenever you like.
MyHermes AI, the default
A new agent starts on "MyHermes AI", MyHermes's own model service, with the model moonshotai/kimi-k2. You don't need a key from anyone:
- It's paid from your AI credits. Instead of a key box, the card shows "balance:" with your credit balance and a "top up" link to SettingsCredits.
- It doesn't accept a key of your own.
Usage and credits explains your balance and how to top it up.
Changing the model
Under "Model", pick from the list, or type any model id your provider offers. While the list loads you see "loading…", then how many models are available, for example "40 available". When there's no list, the line under the box says "Will use:" and the model that will run. If the provider has no model of its own and nothing is typed yet, it says "Pick a model to continue." instead.
ChatGPT (Codex) and Grok (SuperGrok) have no model box here; see Signing in with ChatGPT or Grok.
Press "Save credentials". If you only changed the model, it usually switches without a restart: "Model switched. It takes effect on your next message." That goes for an agent that's asleep too: it stays asleep, and your next message wakes it on the new model. If your agent is in the middle of starting or stopping, or the quick switch doesn't go through, it restarts with the new model instead, and so does any other change.
You can also switch model from the button next to the chat box, without opening Settings. Chat covers that button.
Bring your own key
Providers
The "Provider" grid shows five choices at first (MyHermes AI, OpenRouter, OpenAI, Anthropic and NVIDIA NIM), plus your current one if it's another. "+ 36 more providers" reveals the rest (or "+ 35 more providers" when your current provider is one of the others and already shows). Custom is always last.
| Provider | What you enter |
|---|---|
| MyHermes AI | Nothing: it's paid from your AI credits |
| OpenRouter (tagged "popular") | API key |
| OpenAI | API key |
| ChatGPT (Codex) | Nothing: you sign in with your ChatGPT account (see Signing in with ChatGPT or Grok) |
| Grok (SuperGrok) | Nothing: you sign in with your Grok account (see Signing in with ChatGPT or Grok) |
| Anthropic | API key |
| NVIDIA NIM | API key |
| OpenCode Zen | API key |
| OpenCode Go | API key |
| Ollama | API key, plus a "Base URL", filled in with https://ollama.com/v1 for Ollama Cloud. Change it for a server you run yourself. |
| Groq | API key |
| DeepSeek | API key |
| Google Gemini | API key |
| xAI (API key), Z.AI (GLM), Kimi, Kimi (China), Qwen Cloud, Qwen Coding Plan, MiniMax, MiniMax (China), DeepInfra, Fireworks AI, NovitaAI, Nebius, Hugging Face, Upstage Solar, GMI Cloud, Arcee AI, Vercel AI Gateway, Kilo Code, Xiaomi MiMo, StepFun, Actual Computer, CommandCode, Ramp Router, TokenHub (Tencent), TokenPlan (Tencent), Muse Spark (Meta) | API key |
| Azure Foundry | API key, plus a "Base URL": an https:// address whose host ends in .azure.com, for example https://<resource>.services.ai.azure.com/openai/v1 |
| Custom | A "Base URL" for any OpenAI-compatible endpoint. The key is optional; leave it blank if your endpoint doesn't need one. |
Under each key box there's a link to that provider's site, except for Custom, TokenHub (Tencent) and TokenPlan (Tencent).
Switching provider
- Open SettingsModel and pick the provider. Use "+ more providers" if it isn't shown.
- Paste your key into "API key". For Ollama, Azure Foundry or Custom, check the "Base URL" too.
- Check the model. Picking a provider empties the key box and fills in that provider's usual model (or your saved model, if it's the provider you already use); if it has none, the model box is left empty for you to fill. The list loads once a key is in; you can also type a model id.
- Press "Save credentials".
MyHermes checks your key with the provider before saving anything.
- If the provider rejects it, your old settings stay as they were, and the form shows the provider's reason (sometimes after "LLM credentials not updated:"). Fix the key or the model and save again.
- If it's accepted, your agent restarts on the new provider. The card says your credentials are saved and your agent is starting, with an "Update credentials again →" link. The chat box waits until your agent is back.
- If the card warns that your agent failed to start, your credentials are saved but didn't reach your agent. Press "Update credentials again →" and save once more in a moment.
- If you didn't pick a model for a provider that has none of its own, you see "Pick a model for this provider." Pick or type one and save again. Azure Foundry also refuses a "Base URL" that isn't an
https://address on a host ending in.azure.com.
To go back to MyHermes AI, pick it and save. It needs no key, and the model box fills in with moonshotai/kimi-k2.
MyHermes keeps one model key at a time: switching provider removes the old provider's key, so if you switch back later, paste that key again.
Signing in with ChatGPT or Grok
Instead of pasting a key, you can connect your own ChatGPT or Grok account. Your agent then runs on that account, and its model use goes to that account instead of your AI credits.
- Open SettingsModel (open it), press "+ 36 more providers" (or "+ 35 more providers") and pick "ChatGPT (Codex)" or "Grok (SuperGrok)". Instead of a key box, the card shows "Connect account".
- Press "connect ChatGPT" or "connect Grok". A one-time code appears under "Enter this code on the sign-in page".
- Press "Open sign-in page". It opens in a new tab. Sign in there, and enter the code if the page asks for it ("Copy code" copies it for you).
- Approve the connection, then come back to the MyHermes tab. While the card says "Waiting for sign-in…", stay on the Model card: leaving the Model section or closing Settings stops MyHermes from picking up your approval.
- Once your approval comes through, the card shows "connected to ChatGPT" or "connected to Grok", and MyHermes switches your agent to that account and restarts it.
- There's no "Save credentials" button for these two. Signing in saves your choice.
- Your agent starts on a default model:
gpt-6-lunafor ChatGPT andgrok-4.6for Grok. To change it, use the model button next to the chat box (see Chat). For ChatGPT it lists the models your ChatGPT account offers; for Grok, a fixed set of Grok models. - Signing in replaces your saved key. If you go back to a provider that needs a key, paste its key again. Picking another provider and saving also removes the ChatGPT or Grok sign-in, so coming back means signing in again.
- "reconnect" stays on the card while you're connected. It runs the sign-in again, for example if the chat starts saying your key was rejected.
If something goes wrong, the card says so:
- "Code expired. Try again." The code ran out before your approval came through.
- "Sign-in denied. Try again." Grok reported the sign-in as declined.
- "Sign-in expired or failed. Try again." The sign-in didn't go through.
- An error before any code appears means the sign-in couldn't start.
- A note that you're signed in but your agent couldn't be updated means your sign-in is saved but didn't reach your agent. Sign in again.
To start over, press the card's button. After an error it reads "Try again: ChatGPT" or "Try again: Grok", or still "reconnect" if you were already connected.
Different models for some tasks
Below the model, "▸ Use different models for some tasks (4 roles)" opens four rows. Each one lets a single kind of work run on another model:
| Role | Used when |
|---|---|
| Compression | "Used when conversations get long enough to overflow the context window." |
| Vision | "Used when you send images to your bot." |
| Web extract | "Used when Hermes opens a URL and needs to summarise the page." |
| Session search | "Used for semantic search over your past sessions." |
- A role you leave alone uses your main model. Its tag reads "inherit".
- To change one, tick its "override" box, pick a provider and type a model id (Custom also asks for a base URL). The roles offer MyHermes AI, OpenRouter, OpenAI, Anthropic, NVIDIA NIM, OpenCode Zen, OpenCode Go, Ollama, Groq, DeepSeek, Google Gemini and Custom.
- Keys are shared where they can be. A role on the same provider as your main model says "Reusing primary key.", and two roles on the same provider share one key ("Reusing key from" and the other role's name). Otherwise the row asks for its own key.
- Tags show "saved" for a role you've saved and "modified" for one with unsaved edits.
- Save with "Save credentials", like the rest of the card. That button isn't shown while ChatGPT (Codex) or Grok (SuperGrok) is selected, so set up any roles before you sign in with one of them. A role that says "Reusing primary key." uses your main model's key, and signing in replaces that key, so before you sign in, move that role to another provider with its own key, or turn its override off. Roles with their own key stay when you sign in. To change roles afterwards, pick another provider, paste its key (MyHermes AI needs none) and save: that switches your main model to it, and you'd sign in again to come back.
Your credits and your own key
- On MyHermes AI, what your agent uses is paid from your AI credits.
- On your own key, that provider bills you directly. On a ChatGPT or Grok sign-in, model use goes to that account. Either way your agent's model use doesn't touch your credits, and SettingsUsage says it doesn't track your tokens.
- Your credits stay. The monthly credits that come with your plan keep arriving, and your balance doesn't expire, so it's there if you switch back.
- The built-in Jev skill is billed separately. While your main model is MyHermes AI, Jev is paid from your AI credits. With any other main model, Jev uses an OpenRouter key if your agent has one (as its main model or for one of the four task roles); without one, Jev isn't available. Skills has the details.
On the free tier
Free accounts get SettingsModel too, for the free chat. It shows a shorter list, all at once: MyHermes AI, OpenRouter, OpenAI, Anthropic, NVIDIA NIM, OpenCode Zen, OpenCode Go, Ollama, Groq, DeepSeek, Google Gemini and Custom. Signing in with ChatGPT or Grok, and the rest of the paid list, need a paid plan.
- It starts on MyHermes AI with
myhermes:free, marked "● managed key provisioned". That model is free. Any other MyHermes AI model is charged from your credit balance. - Or bring your own key: pick a provider, paste your key under "API Key" (with a "Get a key from" link to that provider), pick a model and press "Save LLM settings". Once a key is saved, it says "(key already saved)".
Plans and credits covers the free tier's limits.
Tool keys
Open SettingsTool keys (open it). The card is headed "Bring your own search + browser keys". These are keys for the web search and browser services Hermes can use. They're encrypted and only ever handed to your own agent.
| Group | Services |
|---|---|
| "Web search" | Firecrawl, Tavily, Exa, Parallel |
| "Browser automation" | Browserbase (an API key and a Project ID), Browser Use |
- Click "get key ↗" next to a service to get a key from it.
- Paste the key into that service's box. You can fill in several at once.
- Press "Save tool keys".
- "Saved and applied to your agent." means your agent was restarted with the new keys straight away (and woken, if it was asleep).
- If the message says the keys are saved but not that they were applied, they didn't reach your agent just then. Type them in and press "Save tool keys" again in a moment.
- A saved key shows "✓ set" and is never shown again. To replace it, type a new one and save.
- Nothing typed? You see "Enter at least one key to save." If saving fails: "Couldn't save your tool keys. Try again."
Voice keys
Open SettingsVoice (open it). The card is headed "Give your agent a voice." You pick a provider for each half of a spoken conversation and bring your own key, so the provider bills you.
| Half | What it does | Providers |
|---|---|---|
| "Speech → text (STT)" | "Transcribes what you say so your agent can understand it." | OpenAI, Groq (Whisper), Mistral (Voxtral), xAI (Grok) |
| "Text → speech (TTS)" | "Turns your agent's replies into a spoken voice." | OpenAI, ElevenLabs, Google Gemini, MiniMax, Edge (free) |
The xAI entry is only a speech-to-text key. It doesn't connect a Grok account or change your agent's model; signing in with Grok for the model is under SettingsModel. If your main model is "xAI (API key)" and you pick xAI here, or your main model is "MiniMax" and you pick MiniMax here, your agent uses the model's key for voice too.
- Pick a provider in each menu. They start on OpenAI and ElevenLabs until you save something else.
- Paste each provider's key. Its tag reads "needed" until a key is saved, then "configured". Edge needs no key. If you pick the same provider for both halves, one key covers both.
- Press "Save voice settings". You see "✓ Saved. Use the banner at the top to restart your agent and apply the new voice."
- Close Settings and press "Apply changes" in the banner at the top of the dashboard. Memory and background tasks describes that banner.
Voice keys are stored encrypted but aren't checked with the provider when you save, so paste them carefully. When a key is already saved, the box says "Leave blank to keep the saved key."
Which keys are written into your agent's configuration
Every key you save here is encrypted before MyHermes stores it. Your agent receives most keys as environment settings, which backups don't include. The exceptions are worth knowing about, because they do end up in your backups.
- Your main model's key is also written into your agent's configuration file when your main model is on MyHermes AI (the key MyHermes issues for it), OpenAI, NVIDIA NIM, OpenCode Zen, OpenCode Go, Groq, DeepSeek, Google Gemini, Ollama or Custom.
- Other keys aren't written there: not the keys for OpenRouter, Anthropic, Azure Foundry or the providers in the long row of the Providers table, and not the four task roles' keys, tool keys, voice keys or memory keys.
- A ChatGPT or Grok sign-in is kept in a separate file on your agent, and backups don't include it.
- If you use connectors, the same file also holds a key your agent uses to reach them. Connectors explains how that works.
- Backups include that configuration file. So a backup of an agent on one of the providers above contains its model key. Treat your backups as containing secrets. Backups covers what a backup holds.
If something goes wrong
- The form shows the provider's reason (sometimes after "LLM credentials not updated:"). The provider rejected the key or the model. Check both and save again.
- Your agent doesn't answer after switching provider. Check SettingsModel shows the provider and model you expect, then send a message. If it still fails, press "Save credentials" there again without changing anything: it sends your settings to your agent again and restarts it. If you signed in with ChatGPT or Grok, press "reconnect" instead. Chat lists the errors a provider can send back.
- The ChatGPT or Grok sign-in didn't finish. Press the card's button to start again, enter the new code, and stay on the Model card until it says connected. Signing in with ChatGPT or Grok lists each message.
- A ChatGPT model doesn't answer. Pick a model from the model button next to the chat box: it lists the models your ChatGPT account offers.
- Jev stopped working after you switched provider. Jev needs MyHermes AI as your main model or an OpenRouter key. See Skills.
- Voice changes did nothing. They apply only after "Apply changes" in the banner at the top of the dashboard.
What's next
- Skills: what your agent can do, including the built-in Jev skill.
- Usage and credits: your balance, and how to top it up.
- Memory and background tasks: the other settings that apply after a restart.