Which Multi-Model API Provider Should You Connect Hermes Agent To?

The Hermes Agent setup wizard lists GMI Cloud as a built-in provider, labeled "multi-model direct API", so there is no adapter to write: one API key and one base URL give your agent access to closed-source and open-weight models.

Moving between those models is a model ID change, and provider settings stay as they are when a better model ships. GMI's Model-as-a-Service platform hosts LLM, image, video, and audio models behind unified APIs, and the developer endpoint serves 200+ models behind a unified API.

Why a multi-model API fits Hermes

Hermes Agent can run on the chat models in your account that support the tool calling and input types your tasks need, and the model is a setting you control. Single-vendor providers are fine until you want to compare a Claude model with a DeepSeek model on the same recurring task. Connecting each vendor directly means separate keys, separate balances, and separate base URLs.

On GMI's endpoint, that work collapses into one setup:

• OpenAI-compatible endpoint. Hermes talks to https://api.gmi-serving.com/v1. The same key also works for the /v1/responses and /v1/messages routes that GMI documents.
• Closed and open models together. GMI's catalog includes Anthropic, OpenAI, and Google models next to DeepSeek, Qwen, Kimi, GLM, and MiniMax.
• Billing and data handling. Centralized billing covers every model on the key, and zero-retention configurations are available for sensitive workloads.
• Documented for Hermes. GMI's docs include a Hermes Agent tutorial and a second guide for deploying Hermes on Agentbox.

Connect Hermes to GMI Cloud

The Hermes Agent tutorial takes about 10 minutes on a Mac, a Linux machine, or a VPS. You need a GMI Cloud account and, if you want to chat with the agent from your phone, a Telegram account.

1. Install Hermes Agent. Run the one-line installer from the Hermes Agent project. It detects and installs Python, Node, and Git if they are missing. Reload your shell, then run hermes --version to see the installed version.
2. Start the wizard. Run hermes setup and press Enter to begin.
3. Choose GMI Cloud as the provider. Scroll the provider list to "GMI Cloud (multi-model direct API)".
4. Paste your API key. Create one in the GMI console under API keys, name it something like hermes-01, and paste it when prompted.
5. Accept the base URL. Press Enter to keep the default, https://api.gmi-serving.com/v1.
6. Pick a model. The wizard shows the models available to your account. Then type hermes to start a session and send a first message.

The provider step of hermes setup, with GMI Cloud listed as a multi-model direct API. Source: GMI Cloud docs.

From there the wizard continues to a messaging platform. The tutorial uses Telegram: create a bot with BotFather, put TELEGRAM_BOT_TOKEN and TELEGRAM_ALLOWED_USERS in ~/.hermes/.env, and run hermes gateway. Its example task is a standing instruction to find the top three AI news stories every morning and send a summary, which Hermes turns into a scheduled job.

If Hermes returns a 401, rerun hermes setup to re-enter the key, and generate the key after the account has credits. To rotate a key later, edit the GMI_API_KEY value in ~/.hermes/.env.

Switch models without touching the integration

Once Hermes points at GMI, the model is a single setting. You can change it two ways:

hermes model

or permanently in ~/.hermes/config.yaml:

model:
  provider: "gmi"
  default: "deepseek-ai/DeepSeek-V4-Pro"

Model IDs come from three places: the LLM models page in the docs, the model library in the console, or GET /v1/models on your own key. The docs pages for individual models include example requests. Hermes drives tools, so choose chat models that support tool calling.

The default-model step of hermes setup, listing model IDs from several vendors under one GMI account. Source: GMI Cloud docs.

Models to start with

GMI's Hermes tutorial suggests a Claude or DeepSeek model as a strong general-purpose default, and its Agentbox guide uses deepseek-ai/DeepSeek-V4-Pro as the example. Beyond that, the table groups models by the kind of Hermes job you are likely to run.

To settle on a model, give two of them the same batch of recurring tasks and compare how many tool calls each needs, whether the output follows your format, how long it takes, and how many tokens your usage page shows. Switching takes one command, so each run costs a config edit.

Run Hermes on Agentbox

A local Hermes install stops when your laptop sleeps. If you want the agent reachable all day, GMI's Hermes Agent on Agentbox guide moves it to managed compute. Agentbox is in early access, and it provides a runtime for agents with model access and same-network inference behind each call.

The guide's flow, using a Discord bot as the front end:

1. Fork the Hermes Agent repository, add the Agentbox deploy files and publish workflow, and run the workflow in GitHub Actions. It builds the Docker image and publishes it to GitHub Packages; the first build takes 5 to 10 minutes. Make the package public, or keep it private and enable credentials in the Agentbox form with a GitHub token that has read:packages scope.
2. In the console, open Agentbox, choose Register, and enter the image URL. The guide's example uses a container with 2 vCPU and 4 GB of memory and port mapping 443 to 8080.
3. Turn on MaaS integration and select a model. GMI injects GMI_MAAS_API_KEY and GMI_MAAS_BASE_URL for you, so the agent reaches the model API without a key in your image.
4. Add the custom variables from the guide's table (API_SERVER_KEY, DISCORD_BOT_TOKEN, GATEWAY_ALLOW_ALL_USERS, and HERMES_MODEL, plus the optional DISCORD_HOME_CHANNEL), then create an instance. It takes about one to two minutes to come up.

The guide's example sets HERMES_MODEL to deepseek-ai/DeepSeek-V4-Pro and lists any other model as an option. The variable values in the guide are examples for a Discord setup; set the access options that match your own deployment. After registering, Agentbox shows a separate Agentbox API key once. Save it at that moment; it is distinct from your inference key.

Limits and usage tiers

GMI applies rate limits per organization, measured in tokens per minute for LLMs. New organizations start at Tier 1, which is listed at 1,000,000 tokens per minute for LLM models, and the tier rises automatically, within 24 hours, once your total purchases reach the next threshold (voucher redemptions do not count as purchases). The rate limits page lists the current numbers and notes that specific models can be adjusted. An agent that loops on tool calls can use many tokens per minute, so plan your tier around how busy Hermes's schedule will be.

The LLM API reference documents the chat completions endpoint, including streaming, tools, and the per-request parameters.

FAQ

Can credits reload automatically for Hermes's scheduled jobs?

Turn on the Auto Pay toggle on the Account page of the console. While it is on, GMI reloads credits up to once per hour as the balance approaches your auto-pay threshold, charging your default saved card for the configured top-up amount. Scheduled work such as the morning news brief draws on that same credit balance while hermes gateway runs.

Can I see what Hermes spends separately from my other apps on the same GMI account?

Give Hermes its own key, such as the hermes-01 from the setup steps, and let your other apps use theirs. In Billing, open Serverless Usage and pick hermes-01 in the API Key filter. Adding Model Name breaks it down further, and Hourly granularity with Local Timezone brings out the spikes from scheduled jobs.

Run one task, then change the model ID

Give Hermes the task you care about on the default model from the table, then change the model ID and run it again. The docs cover the full model list and API reference, and GMI Cloud sales can help with volume pricing, custom limits, or a dedicated setup.