Run Hermes Agent: A Self-Evolving AI You Own Locally

Step 1: Clone the Repository

Open a terminal and pull down the project:

bash
git clone https://github.com/NousResearch/hermes-agent
cd hermes-agent

No account required. Everything runs locally.

Step 2: Install Dependencies

Hermes Agent uses Python. Create a virtual environment and install:

bash
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

This pulls in DSPy, the optimization layer that powers self-evolution, along with all channel adapters.

Step 3: Configure Your Model

Copy the example config and point Hermes at your preferred model:

bash
cp .env.example .env

Open .env and set your model endpoint. Hermes accepts OpenAI-compatible APIs, so this covers Claude (via proxy), GPT-4o, a local Ollama instance running Llama or Mistral, or any other compatible backend. If you are running locally with Ollama:

MODEL_PROVIDER=ollama
MODEL_NAME=llama3

For cloud models, drop in your API key for the relevant provider.

Step 4: Choose Your Deploy Channel

Hermes supports multiple interfaces. Pick the one that fits your workflow:

• CLI -- fastest to start, no tokens required: python main.py --channel cli
• Telegram -- create a bot via BotFather, paste the token into .env
• Discord / Slack / WhatsApp / Signal -- each has a dedicated adapter section in .env with instructions inline

For a first run, CLI is recommended so you can watch the prompt-evolution logs directly in your terminal.

Step 5: Run the Agent and Watch It Evolve

Start a session:

bash
python main.py --channel cli

Have a real conversation -- ask it to help with something you do repeatedly. After the session ends, Hermes uses DSPy + GEPA to score which reasoning steps produced good outcomes and rewrites its internal prompts before your next session. Check the /prompts directory to see exactly what changed.

First prompt worth running: Ask it to help you draft a recurring work deliverable (a weekly report, a code review checklist, a client update). Do it three sessions in a row. By session three, you should see it skip the setup questions it asked in session one.

---

Why This Works

DSPy treats prompts as optimizable programs rather than static strings, and GEPA provides a gradient-like signal from real conversational outcomes. Because the optimization happens between sessions on your own machine, the improvements are compounding -- each conversation becomes training data for the next prompt revision -- without sending anything to a third-party trainer.

---

Pro Tips

• Commit your /prompts folder to git. You get a full version history of how your agent evolved and can roll back any change that made things worse.
• Use the same task type repeatedly early on. Hermes needs enough signal to spot patterns. Varied one-off tasks slow the optimization loop.
• Swap models without losing progress. Evolved prompts are model-agnostic text files. Switch from GPT-4o to a local Mistral and your optimized prompts carry over.
• Pin the DSPy version in requirements.txt. DSPy is under active development and minor version bumps can change optimization behavior in ways that are hard to debug.

---

The full source, adapter docs, and contribution guide live at github.com/NousResearch/hermes-agent. Star it if you want to track the self-evolution roadmap -- the project added 30K stars in a single week and the issue tracker is moving fast.