First Reliable Run

The first Hermes run should not be a large task. Its job is to prove the smallest chain: the command starts, provider auth works, the model responds, the working directory is correct, output is bounded, and you know where to look if it fails.

Start from the working directory

Hermes reads the current directory, project context, and available tools. Start from a low-risk project, ideally one you can inspect with Git:

cd /path/to/project
git status --short
hermes

For a trial run, a temporary directory with a README is enough. Do not start the first session from a production repository, a secrets directory, or a working tree full of important uncommitted changes.

Configure one model first

The official quickstart treats hermes model as the key setup step. It walks through provider, model, and authentication choices:

hermes model

Do not begin with routing or fallback. Get one provider and one model working first. Compare quality, cost, latency, and context length after the basic path is stable.

If you use an OpenAI-compatible local server or self-hosted model, test the endpoint, model name, and key with the smallest possible request first. Many “agent” problems are really model-route problems.

Make the first task read-only

Use a prompt like:

Read README.md and summarize what this project does in five bullets. Do not modify files.

This checks three things: Hermes sees the right directory, the model can return a structured answer, and the session respects a no-write boundary.

If this is unstable, do not add browser tools, shell commands, gateway, or cron yet. Go back to the install and provider layers.

Then write only to a temporary path

After the read-only task works, try a low-risk write:

Read README.md and write a short summary to /tmp/hermes-first-run.md. Do not modify project files.

Then inspect the result yourself:

ls -l /tmp/hermes-first-run.md
sed -n '1,80p' /tmp/hermes-first-run.md
git status --short

git status --short should not show project changes caused by the task. If it does, inspect the diff before expanding scope.

Record the first successful path

A useful team note can be very short:

OS:
Shell:
Hermes version:
Python version:
Provider:
Model:
First successful command:
Working directory:

This is not paperwork for its own sake. It tells the next person which version, model, and environment made the behavior true.

If the first run fails

Debug one layer at a time:

Symptom	First place to look
`hermes` is unavailable	Installation
Provider auth fails	Key source, account permission, model name, region.
Output is empty, slow, or messy	Model choice, context length, prompt boundary.
README cannot be read	Current working directory or file permissions.
Output is written to the wrong place	Prompt boundary, tool permissions, working directory.

When to move on

Once you can complete a read-only task and a temporary write, and you can name the provider, model, working directory, and state path, continue with:

First Reliable Run

First Reliable Run

Start from the working directory

Configure one model first

Make the first task read-only

Then write only to a temporary path

Record the first successful path

If the first run fails

When to move on

References