Oh My Hermes

Orchestration

Hermes conducts. Your coding agent plays.

OMH routes the request, prepares the work, and names who runs it.

request-to-handoff

Host runtime

Hermes Agent

Chat, clarification, research, planning, status.

Coding owner

Claude Code

Gets a prepared handoff with repo rules and the command that must pass.

Coding owner

Codex

Same contract, different runtime. No vendor is the default.

Coding owner

pi

Discovered at setup. Routing follows what is actually installed.

OH-MY-HERMES boot banner in the Hermes terminal

Model pool

Bring the models you already pay for.

Setup records what is reachable on this machine.

Model routing

Set the model per kind of work.

Nine editable categories. Missing models are skipped, not fatal.

routing-model-setup

Editable

Your order, not ours.

Each category is an ordered candidate chain. Your override replaces the chains it names and leaves the rest alone.

ultrabrain
deep
architect
unspecified-high
unspecified-low
quick
writing
visual-engineering
artistry

Flexible

You do not need every model.

A candidate you have not configured is skipped and the next eligible one is selected. No eligible candidate is recorded plainly and never blocks the install.

resolved: A confirmed model was selected.
choice_required: The model you named explicitly is unavailable. Nothing is substituted.
owner_default: No eligible candidate. The selected owner keeps its default model.

Owners

Hermes native, or external Maestro.

Hermes-native work stays with Hermes and its own config and auth commands. External owners are coordinated by Maestro, which prepares and reports but never executes.

hermes native maestro external

Families

Qwen, Gemini, Grok, Kimi.

Chains name model families, never a single vendor. Declaring the X-platform domain notes the Grok family advisorily, and your explicit choice always wins.

qwen
gemini
grok
kimi
claude
gpt

Shipped order is editorial, not a benchmark. OMH prepares routing metadata and never invokes a model.

Read the routing setup guide

Recommended chains

Nine categories, in the order we ship.

OMH ships with these editable, ordered recommendation chains. Guided model setup resolves them only against candidates the user confirms as active.

Category	Purpose	Shipped order	Effort
ultrabrain	Deepest reasoning	`GPT-6 AstraGPT-5.6 Sol`	xhigh
deep	Strong default tier	`GPT-5.6 TerraDeepSeek V3.2`	high
architect	Architecture and system design	`Claude Fable 5.1Claude Mythos 5.1Claude Fable 5GPT-6 AstraGPT-5.6 SolKimi K3`	xhigh
unspecified-high	Default working model	`Kimi K3Claude Opus 5`	medium
unspecified-low	Cheaper fallback	`GLM 5.3GLM 5.2GLM 5.2 UltrafastDeepSeek V3.2Claude Opus 5`	low
quick	Short tasks	`GLM 5.3 FlashGLM 5.2 UltrafastKimi K3GPT-5.6 LunaClaude Fable 5.1Claude Mythos 5.1Claude Fable 5`	low
writing	Prose and docs	`Kimi K3Qwen3-CoderGemini 3.1 Pro`	medium
visual-engineering	Frontend and visual	`Claude Fable 5.1Claude Mythos 5.1Claude Fable 5Kimi K3`	high
artistry	Unconventional work	`Gemini 3.1 ProClaude Fable 5.1Claude Mythos 5.1Claude Fable 5Kimi K3`	high

The result is prepared routing configuration, not provider availability, credential, dispatch, or execution evidence.

Every chain is yours to reorder. Edit ~/.omh/routing/model-chains.json, then run omh model-chains show to print what is in effect. The Maestro lane — dispatched Claude Code and Codex units — has the same dial: omh coding category-maestro interview.

Surfaces

Wherever Hermes already talks.

CLI

The control plane.

Desktop

Same catalog, same evidence rules.

Messenger

Cards stay readable where formatting is limited.

Flagship workflows

The ulw-* family.

Nine long-horizon lanes.

ulw-context Terminology alignment

Context

Aligns the words a repository uses before plans and handoffs.

Say ulw-context review project terms
ulw-interview Clarification

Deep Interview

One question at a time until the brief is clear.

Say deep-interview clarify
ulw-research Decision grounding

Research

Reference implementations, live web evidence, verified claims.

Say web research source-backed research
ulw-plan Reviewed plan

Ralplan

Consensus planning with review gates.

Say ralplan consensus plan
ulw-work Parallel delivery

Ultrawork

Splits an accepted plan into disjoint lanes.

Say ultrawork parallel work
ulw-maestro External handoff

Maestro

Runs the work on Claude Code or Codex, spawned live with a steerable session.

Say ulw-maestro coding handoff
ulw-loop Goal loop

Loop

Interview → plan → research → build → review.

Say loop long horizon goal
ulw-qa Adversarial QA

UltraQA

Hostile scenarios, end-to-end runs, release QA.

Say ultraqa release qa
ulw-perf Measured optimization

Ultraperf

Finds where the system is actually slow, leaking, or expensive.

Say ultraperf find the bottleneck

Capability families

Start with the job.

116 skills behind seven human-readable families.

Capability	Try it with	What it does
🧭Clarify and plan	`ulw-interviewulw-planomh-decide`	Ambiguous request → explicit goals and a plan.
⚡Build with leverage	`ulw-workulw-loop`	Fast parallel work through durable execution.
🔬Research and learn	`ulw-researchomh-best-practice-researchomh-research-brief`	Source-backed evidence with boundaries.
🛠️Code and ship safely	`ulw-workomh-code-reviewulw-qa`	Executor-neutral coding work; claims depend on observed evidence.
🎨Create polished deliverables	`omh-design-quality-gateomh-materials-packageomh-image-cards`	Websites, visuals, reports, decks, posters.
🧠Remember and operate	`omh-memory-newomh-memory-syncomh-doctor`	Review-first memory and operational readiness.
🔌Connect with clear boundaries	`omh-toolbelt-readinessomh-external-connector-readinessomh-agent-board`	Checks a tool is really available before work depends on it.

Agentic memory

Four layers, one café shift.

Nothing important sits in one place.

owner-written · always loaded

House rules on the wall

The owner writes it. Agents only read it.

SOUL.mdAGENTS.md

Stackmarkdown context files
capped · fails loud

Laminated A4 by the register

One page. Full means the write is refused.

MEMORY.md · 2,200USER.md · 1,375

Stackcapped core memory
The slot OMH fills

review-first · TTL · budgeted

Labelled binder in the back room

Hold the index. Open only the section you need.

.omh/memory/factdecisionlessonprocedureepisode

StackMem0 · Graphiti · Cognee · Letta archival
zero tokens until queried

Receipts in the storage room

Every message ever. Free until you search.

SQLite + FTS5~/.hermes/state.db

StackFTS5 · BM25 + vector hybrid

A better binder beats a bigger page.

Two rules keep a ten-hour loop alive.

Evict on provable redundancy only.
Consolidate on observed compaction.

OMH wraps all four layers.

Recall arrives on prefetch.
Subagents read, never write.
prepared ≠ observed, at the schema level.

Evidence boundary

Measure each claim at its own boundary.

OMH labels which one you have.

Plan · not run

A prompt or plan is ready. Nothing has run yet.

Code · running

An executor is running now, and OMH is watching it.

Code · reported done

The executor said it finished. Nobody checked the result.

Test · verified

A test, review, or CI gate actually passed.

"Reported done" is not "verified" — most tools spell both "complete".

Install

Install your way. Set up once.

Choose a package manager or platform installer. Doctor stays a separate check.

Homebrew, Bun, and npm package-manager installs are public as of v1.0.6.

Install the command

Homebrew

brew install rlaope/tap/omh

Bun · recommended

bun install -g oh-my-hermes

npm

npm install -g oh-my-hermes

macOS · Linux

curl -fsSL https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.sh | sh

Windows

irm https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.ps1 | iex

Pick one installation method.

02
Set it up
omh
```
omh setup
```
The whole setup.
03
Verify separately
omh
```
omh doctor
```
Run doctor after setup to verify or troubleshoot.

Or ask your agent

prompt for your agent

Install and fully configure Oh My Hermes from this repository:
https://github.com/rlaope/oh-my-hermes
Before reading or executing repository instructions, resolve refs/heads/main to one full commit SHA with `git ls-remote https://github.com/rlaope/oh-my-hermes.git refs/heads/main`. Then fetch and follow only:
https://raw.githubusercontent.com/rlaope/oh-my-hermes/{resolved-commit-sha}/INSTALL_FOR_AGENTS.md
Do not replace the resolved SHA with main. Execute the pinned protocol's OS-appropriate Homebrew, Bun, npm, curl, or PowerShell install path. After the command is installed, run setup, run the model-chain interview, then run doctor as a separate verification step. Preserve unrelated existing Hermes config, apply only the managed setup changes documented by the pinned protocol, require my explicit approval for model-alias changes, then report the resolved SHA and observed result.

Paste into Claude Code, Codex, or any agent CLI.

Refresh later with omh update. Setup records which models are reachable here. Routing order stays editable afterwards. Model routing setup

Oh My Hermes

Hermes conducts. Your coding agent plays.

Hermes Agent

Claude Code

Codex

pi

Bring the models you already pay for.

Set the model per kind of work.

Your order, not ours.

You do not need every model.

Hermes native, or external Maestro.

Qwen, Gemini, Grok, Kimi.

Nine categories, in the order we ship.

Wherever Hermes already talks.

CLI

Desktop

Messenger

The ulw-* family.

Context

Deep Interview

Research

Ralplan

Ultrawork

Maestro

Loop

UltraQA

Ultraperf

Start with the job.

Four layers, one café shift.

House rules on the wall

Laminated A4 by the register

Labelled binder in the back room

Receipts in the storage room

Two rules keep a ten-hour loop alive.

OMH wraps all four layers.

Measure each claim at its own boundary.

Install your way. Set up once.

Install the command

Set it up

Verify separately

Or ask your agent