Hermes Agent + DeepSeek on the Gonka network — an autonomous agent for pennies

Installing Hermes (Linux, macOS, WSL2). The official installer will automatically pull Python, Node.js, ripgrep and ffmpeg:

# Linux / macOS / WSL2
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

# Windows (PowerShell)
iex (irm https://hermes-agent.nousresearch.com/install.ps1)

After installation, restart your shell and check if it runs:

source ~/.bashrc   # or ~/.zshrc
hermes             # interactive CLI

JoinGonka API-key: if you don't have one yet, register at gate.joingonka.ai/register, verify your email address — 3M free tokens will be credited to your account — and create a key with the jg- prefix in the Dashboard. One key and one balance work for all models immediately.

Hermes works with any OpenAI-compatible endpoint: if the server responds to /v1/chat/completions, Hermes can be pointed to it. The JoinGonka Gateway is exactly that. There are three ways to configure it.

Method 1: our installer (one command, recommended). It will add the provider to your Hermes config, set DeepSeek V4 Flash as the default model, and perform a live check to verify the gateway accepts the key and model:

npx @joingonka/setup --tool hermes

Method 2: setup wizard. Run this in your terminal outside of an active session:

hermes model

In the provider list, select «Custom endpoint (self-hosted / VLLM / etc.)» and enter the three values:

• API base URL: https://gate.joingonka.ai/v1
• API key: jg-your-key
• Model name: deepseek-ai/DeepSeek-V4-Flash-0731

Method 3: manual config edit. Hermes stores model settings in ~/.hermes/config.yaml — this is the single source of truth. Add this section:

# ~/.hermes/config.yaml
model:
  provider: custom
  model: deepseek-ai/DeepSeek-V4-Flash-0731
  base_url: https://gate.joingonka.ai/v1
  api_key: jg-your-key

⚠️ Key placement — model.api_key only. Hermes documentation allows OPENAI_API_KEY in ~/.hermes/.env, but this does not work for custom endpoints: Hermes reads the key only from model.api_key, and if you use the key in .env, the gateway will return 401 Invalid API key. Verified on Hermes v0.20.4.

Verification: run hermes and write "Create a hello.py file with a hello world function". The agent should create the file and show the result. You can check the current configuration state using hermes config show | grep '^model\.' and hermes status.

Switching on the fly: inside a session, use the command /model custom:deepseek-ai/DeepSeek-V4-Flash-0731. And /model custom without a name will automatically fetch the model from the endpoint if there is only one.

Hermes is an agentic tool with memory and skills. It doesn't just respond with one line: it reads files, writes code, runs commands, compresses context, and performs long background tasks via a scheduler. Each action is an API call, and tens of millions of tokens accumulate over a day of autonomous work. Let's compare the costs of typical sessions:

With JoinGonka Gateway, Hermes can be kept running 24/7: nightly reports, weekly audits, background tasks via a cron-scheduler — without worrying about the bill. At Anthropic prices, each such run would have to be calculated manually. The starting balance upon registration is enough to run the agent on a real task and see your own consumption before your first top-up.

Through the Gateway, three Gonka models are immediately available to the user — all of them support native tool calling (OpenAI function calling), so Hermes skills and tools work reliably without parsing text responses:

About token limits in Hermes. In config.yaml, the context_length field is the total context window (input + output), while max_tokens is the ceiling only for the response length. It is better not to set either: Hermes will determine the correct values from the provider itself. If you set max_tokens manually, keep it within the model limits: via Gateway, this is 32768 for DeepSeek V4 Flash and 8192 for MiniMax M2.7 and GLM-5.3 Flash. You can change the model at any time with the hermes model command or directly in a session via /model.

Similar terminal tools in our knowledge base: Claude Code, Cline.