rlaope/oh-my-hermes

oh-my-hermes (OMH) is an add-on for Hermes Agent, the coding agent from Nous Research. It sits between your request and Hermes and decides which model and reasoning effort each piece of work gets, loads specialist skills for it, splits the work into parallel lanes, and reports what was actually checked versus merely attempted. It does not patch or replace Hermes, so your existing agent workflow stays intact README.

What changes in the workflow

The headline path is /dag-ulw: a goal becomes a graph of small nodes, nodes with no dependencies run at once, and a separate node verifies the goal before a report is produced. Lanes that run concurrently are kept from sharing a write scope, and failed nodes are repaired before reporting. Ten ulw- workflows cover the other stages, from interviewing you about an unclear idea through planning, parallel execution, review loops and performance measurement workflow reference.

Two mechanisms carry most of the practical difference. First, each lane is tagged with a work category such as ultrabrain, deep, visual-engineering or quick, and that category selects a model chain and reasoning effort; if a provider rejects a model, the chain moves to the next rather than silently downgrading, and chains are editable. Second, statuses separate stage from confidence: Plan · not run, Code · reported done and Test · verified are distinct, and only a passed test, review or CI gate counts as verified evidence rules. This matters for anyone who has been burned by an agent saying "done" without anything being checked.

Requirements and install paths

The main install is a shell script for macOS/Linux or PowerShell 5.1+ on Windows, followed by omh setup. Homebrew, Bun and npm are alternatives; Bun and npm require Node 18+ and Python 3.11+ on PATH, while Homebrew brings its own Python. Agent-driven and Hermes skill-tap install paths also exist installation docs. OMH is built around Hermes, so it is not a standalone coding assistant.

Updates run through omh update, which routes through the owning package manager, and omh doctor verifies or troubleshoots the setup. Removing the command package preserves OMH state; full removal requires omh uninstall --all first. Installing under a different method than the original is handled through documented reconciliation, and the agent-driven install pins an exact commit SHA before fetching instructions installation docs.

Documented limits

Most capability claims are the project's own. The comparison of Hermes alone versus Hermes through OMH has its corpus and offline pilot ready, but no measured run is published yet product A/B. The source also reports no runtime validation for some static reasoning about code. Treat feature descriptions as author statements rather than independently reproduced results.

Licence

MIT repository.