oh-my-hermes

Oh My Hermes (OMH) is a command‑line tool that connects the Hermes Agent to your development environment. It provides a set of predefined model chains, routing rules, and status tracking so you can run AI‑assisted coding tasks directly from the IDE.

The core idea is to treat each kind of work, such as architecture design, quick prose, or long‑horizon research, as a separate lane. Every lane is tagged with a category, and the tool selects a chain of models (for example Claude Opus 5.5 → GPT‑6.1 Sol → Kimi K3.1 Pro) based on that category and a configurable “effort” level. You can edit the routing file (`~/.omh/routing/model-chains.json`) to reorder models or change the effort, and the picker UI lets you switch categories with arrow keys and adjust effort with “‑” and “+”.

To start, you install OMH with a curl script, then run `omh setup` to configure the environment. The `omh model setup` command presents a list of supported providers, including OpenAI, Claude, Gemini, Grok, DeepSeek, and others, so you can add credentials for the models you want to use. After the setup, you invoke a workflow such as `ulw-work` or `dag-ulw` to launch a parallel or DAG‑based run. The tool shows a textual UI with multiple lanes and layers, indicating the current stage (Plan, Code running, Code reported done, Test verified). Evidence is recorded at each step, and if a provider rejects a model the chain falls back to the next one without silent downgrade.

The system also includes “flagship workflows” like `ulw-context` for terminology alignment, `ulw-interview` for incremental clarification, and `ulw-research` for web‑based evidence gathering. Each workflow follows a predefined sequence of lanes, and you can observe the live execution in the terminal UI, where events such as “streaming lane events” and “fallback” are displayed. The routing configuration is separate from provider availability, so you need to ensure credentials are valid before a run.

Key features

• Model routing per lane with editable JSON chains
• Support for 18 work categories and multiple effort levels
• Parallel and DAG execution of lanes via `ulw-work` and `dag-ulw`
• Real‑time terminal UI showing lane status and evidence checkpoints
• Integrated setup for many LLM providers (OpenAI, Claude, Gemini, etc.)
• Fallback handling when a model is unavailable or rejected