Hermes Agent setup: install and build self-improving skills

What Hermes does differently

Most AI agent frameworks treat every task as a fresh reasoning problem. Hermes takes a different approach: after completing a task, it saves the successful execution path as a skill file. The next time it encounters a similar request, it loads the saved skill and follows the proven steps. Over time, the agent builds a library of playbooks tuned to your specific workflows.

Skills are plain-text YAML files stored in a local directory. You can read, edit, or delete them. There is nothing opaque about the system. If a skill produces bad results, you delete it and Hermes falls back to reasoning from scratch on the next matching request.

Install Hermes

Hermes ships as a single binary. Install it with one command:

curl -fsSL https://hermes-agent.dev/install.sh | bash

This downloads the latest release for your platform (macOS arm64, macOS x64, or Linux x64) and places the binary in ~/.local/bin/hermes. Verify the install:

hermes --version

If your shell does not find the command, add ~/.local/bin to your PATH. The installer prints instructions for this if it detects the directory is not in your PATH.

Configure the LLM provider

Generate a default config file:

hermes init

This creates ~/.config/hermes/config.yaml. Open it and set your provider:

llm:
  provider: anthropic
  model: claude-sonnet-4-20250514
  api_key: sk-ant-your-key-here

Hermes also supports OpenAI and Ollama. For OpenAI:

llm:
  provider: openai
  model: gpt-4o
  api_key: sk-your-key-here

For local Ollama:

llm:
  provider: ollama
  model: llama3.1:70b
  base_url: http://localhost:11434

Connect Telegram

Set up a Telegram bot to send tasks from your phone:

1. Message @BotFather on Telegram and send /newbot.
2. Choose a name and username for the bot. BotFather returns a token.
3. Add the token to your config:

channels:
  telegram:
    enabled: true
    token: "7123456789:AAF..."
    allowed_users:
      - 123456789  # your Telegram user ID

Get your user ID by messaging @userinfobot. The allowed_users whitelist prevents anyone else from using your agent.

Run your first task

Start Hermes:

hermes start

Open your Telegram bot and send a task:

Write a bash script that checks disk usage on all mounted volumes and flags any over 80%

Hermes plans the task, executes it, and sends the result back. The first run on a new task type triggers skill generation. You will see a log line like:

skill saved: ~/.config/hermes/skills/disk-usage-check.yaml

How skills work

After a successful task, Hermes analyzes the steps it took and writes a skill file. A skill file looks like this:

name: disk-usage-check
trigger: "check disk usage"
steps:
  - action: shell
    command: "df -h --output=pcent,target | tail -n +2"
  - action: parse
    pattern: "extract lines where usage > threshold"
  - action: respond
    template: "Volumes over {threshold}%: {results}"
tags: [monitoring, disk, bash]

When the next request matches the trigger pattern, Hermes loads this skill instead of reasoning from scratch. The result is faster execution and more consistent output. Skills accumulate over time, and the agent gets better at your specific recurring tasks.

To review all saved skills:

hermes skills list

To delete a skill that is not working well:

hermes skills delete disk-usage-check

Deploy on Modal for always-on operation

Running Hermes on your laptop means the agent stops when your machine sleeps. For persistent operation, deploy to Modal, a serverless compute platform that bills per second of active execution.

1. Install the Modal CLI: pip install modal
2. Authenticate: modal setup
3. Deploy Hermes:

hermes deploy modal \
  --llm-provider anthropic \
  --llm-key sk-ant-your-key \
  --telegram-token 7123456789:AAF...

The deploy command packages Hermes into a Modal function that wakes up when a Telegram message arrives and goes to sleep after processing. Idle time costs nothing. Active processing costs roughly $0.0001 per second on Modal's standard GPU-free tier.

For a typical personal agent handling 20 to 30 tasks per day, this works out to $2 to $5 per month in compute costs, plus your LLM API spend.

Troubleshooting

• hermes: command not found: Add ~/.local/bin to your PATH. Run export PATH="$HOME/.local/bin:$PATH" and add it to your shell profile.
• Skills not generating: Skill generation only triggers on tasks that complete successfully. Check hermes logs for errors during execution.
• Telegram messages not arriving: Confirm the bot token is correct and the user ID is in the allowed list. Restart Hermes after config changes.
• Modal deploy fails: Run modal setup to re-authenticate. Check that your Modal account has billing enabled (required even for the free tier).

Next steps

With Hermes running and skills accumulating, explore connecting additional channels (Discord, Slack), adding tool integrations for GitHub or database access, and tuning the skill trigger patterns for higher match accuracy. See the agentic AI overview for how Hermes fits into a multi-agent workflow.

Frequently asked questions