Oh My Hermes

Oh My Hermes Start Guide

OMH is a layer on top of Hermes Agent. You keep talking to Hermes; OMH splits the work into lanes and runs each one on the model that suits it, at the same time.

It never patches Hermes, and it never calls planned work done.

$curl -fsSL https://omher.app/install | sh
Install
01

What OMH does

Models by category

Every lane gets a category, like quick or deep. The category picks the model and how hard it thinks.

Parallel runs with dag-ulw

Type /dag-ulw and a goal. It becomes small steps, the ready ones run in parallel, and a separate step checks the result.

Expert skills

A hundred-plus specialists: frontend, backend, Rust, security review. You never call them; the router brings the right ones in.

Skills →

Long-term memory

Nothing is remembered silently: each candidate is kept, refused or deferred with a written reason. The next session gets the memories that fit its task, cut to a token budget.

Memory ↗
02

Install

1

Install omh

macOS and Linux. It finds Python 3.11+ on its own.

$curl -fsSL https://omher.app/install | sh
2

Connect it to Hermes

Installs the skills and registers OMH with Hermes.

$omh setup
3

Pick your models

Opens the picker, one row per category. The arrow keys change the model and the effort.

$omh model
Windows, or Hermes + OMH + the desktop app in one go

Windows (PowerShell)

$irm https://omher.app/install.ps1 | iex

Hermes if it is missing, then OMH, then the omher app (beta)

$curl -fsSL https://omher.app/omher | sh
Every install option ↗
03

Categories and model routing

A category is a kind of work. Each one keeps its models in order: first choice, then the next.

quickShort tasks
  1. 1GLM 5.3 Flashlow
  2. 2GPT-6 Lunalow
  3. 3Claude Haiku 5.5low
  4. 4Kimi K3.1 Prolow
deepStrong default
  1. 1Claude Opus 5.5high
  2. 2GPT-6.2 Solhigh
  3. 3Kimi K3.1 Prohigh
visual-engineeringScreens and frontend
  1. 1Claude Fable 5.2high
  2. 2Claude Opus 5.5high
  3. 3Gemini 4 Argonhigh
  4. 4Kimi K3.1 Prohigh
deep
Claude Opus 5.5highProvider said no
GPT-6.2 SolhighNext in line runs

If a provider refuses a model, the next one in the list takes over. OMH never quietly swaps in something weaker.

The models shown are today's defaults. Run omh model to see yours; every list is yours to edit.

Model routing in detail →
04

plan and dag-ulw

planulw-plan · ulw-work

ulw-plan writes the list and gets it agreed. ulw-work then works through it.

  1. Read the code
  2. Reproduce
  3. Fix
  4. Write a test

dag-ulw

dag-ulw turns the list into a flowchart. Only a step that needs an earlier one waits; the rest run at the same time. A separate step checks the goal, and anything that failed is fixed before you get the report.

same time if the check fails, fix again waitingRead the codewaitingReproducewaitingFixwaitingWrite a testwaitingCheck the goal same time if the check fails, fix again waitingRead the codewaitingReproducewaitingFixwaitingWrite a testwaitingCheck the goal

Not sure which one? Type /dag-ulw and what you want.

05

Updates and compatibility checks

When a new model or a new Hermes ships, OMH does not just follow. It checks that things still work, then updates.

New model

OMH first checks that its router recognizes the model and which effort levels the provider accepts. Only then does it take a place in the category lists.

How it is done ↗

New Hermes

Every night the OMH plugin is loaded into a real Hermes, both the stable release and main. If any check is missing, the run fails.

The nightly check ↗

Skill sources

A skill shaped by outside work records the upstream commit it was reviewed against. When upstream changes, the change is folded in or turned down, and the decision is written down; CI blocks a record without it.

The source registry ↗

All you run is omh update.

$omh update
06

The omh TUI

Bare omh opens OMH's own terminal on your Hermes. Every model call and approval is still Hermes'; the screen is OMH's. /dag-ulw draws the graph, then a live tree with each lane's model, effort and tokens.

  • /themeChange the launch look
  • ! lsRun one shell line
  • Esc EscRewind
  • omh --resumeCome back to a session
07

omher desktop (beta)

ADE means agent development environment. Your Hermes bots sit on the left, each with its work threads; the chat is in the middle; terminals open in a panel at the bottom with Ctrl `, as in VS Code.

The omher desktop app: the bot Kami with three work threads, the bots Friren and Juno, terminals for a shell, Codex and Claude Code, and Kami's chat in the middle

One command installs Hermes if it is missing, OMH, and the app.

$curl -fsSL https://omher.app/omher | sh
Download app
08

More features

ulw-maestro

Hands a coding task to Claude Code or Codex, and lets you steer it while it runs.

Verification status

A run counts as verified only when a test, a review or CI actually passed.

LSP integration

Definitions, references and renames come from the language server, not text search. Off by default.