Oh My Hermes (OMH)
A professional operating layer for Hermes Agent: per-model routing, parallel coding, 108 specialist skills, and reviewable long-term memory — install once, keep Hermes.
Evidence shows well-designed permission boundaries: SECURITY.md commits to not silently patching Hermes internals, least-privilege CI permissions, install --dry-run, toolcall rules, and approval tiers. Deductions: these are largely declarations; the actual installer script bodies are not in the provided files, model-route data flow depends on user config with only an assertion that credentials are not stored, and no third-party audit exists. The reviewer-approved memory writes are good design, but approval mechanics are not shown.
Strong self-consistency: multiple generated-doc drift gates (docs workflows/roles/ulw --check) enforce doc-code alignment; zero runtime dependencies with multi-channel install cross-validated; failure semantics are explicitly layered (reported done ≠ verified, cost 'unknown' not $0, refusals of silent downgrades). Full marks are justified because CI and test files themselves enforce these constraints.
Broad scenario coverage (multilingual READMEs, macOS/Linux/Windows, dedicated PowerShell 5.1 parse tests). Deductions: capability boundaries and trigger precision rest mainly on documented claims (108 skills, routing corpus, negative cases); the provided files show the framework but not the corpus quality itself.
Excellent information architecture (docs map, layered INSTALLATION/CAPABILITIES/WORKFLOWS/ARCHITECTURE); install notes cover five channels plus uninstall/rollback; complete MIT license; maintenance responsibility evidenced by SECURITY.md, issue numbers, and named policy owners. Deductions: no standalone CHANGELOG file in evidence, naming stability only implied by a 'never OMO' test suggesting a past rename, and known limitations are covered only by a brief 'early' note.
Output usability is supported by HUD state tiering and evidence-state design; marginal value proposition is clear (model routing, long-term memory, parallel fanout). Deductions: headline effect claims ($4.29→$0.66, 23→5 minutes) are statically unverifiable numbers; cost-benefit has anti-fabrication gates inside the benchmark framework, but no reviewable raw run data appears in the provided files.
Fact/inference separation is a highlight: the repo explicitly distinguishes preference vs benchmark, reported done vs verified, prepared config vs provider availability. Deductions: key performance figures lack traceable raw run data, and cross-source corroboration covers mechanisms (CI, schemas, tests) but not effectiveness claims, which remain self-reported.
- The installer is recommended via curl | sh; static review could not inspect the script body. Download and review before executing.
- Cost/latency comparison figures in the README (e.g., $4.29→$0.66) are self-reported and not statically verifiable; do not treat them as procurement evidence.
- Publisher identity is unverified; audit the installer's writes to Hermes config and PATH before enterprise adoption.
- Capability claims (108 skills, routing corpus) lack a raw manifest in evidence; trigger quality must be validated by an actual install.
What does this agent do, and when should you use it?
oh-my-hermes is an MIT-licensed plugin for NousResearch's Hermes Agent (GitHub: rlaope/oh-my-hermes) that positions itself as an operating layer above Hermes-native skills, not a replacement. It routes every request into one of nine editable model categories (ultrabrain, deep, quick, etc.), each an ordered chain of model + reasoning effort that falls through when a provider rejects a model. The coding side adds per-family calibrated prompting across 13 model families, the ulw-work parallel workflow (disjoint file ownership, worktrees branched from a pinned SHA, typed results), and 108 omh-* specialist skills injected into runs automatically. It ships a HUD terminal surface with per-lane cost/turn/token rows and a phase todo, plus a review-gated, file-backed long-term memory that never touches Hermes' own memory. Its core discipline is evidence before claims: an executor saying it finished is only reported done; verified appears only after a test, review, or CI gate passes.
OMH installs and configures through the omh CLI (omh setup interviews your providers and reorders model chains, omh update upgrades, omh doctor verifies). At runtime it scores each request before dispatch, selects a category chain, and rejects a dispatch that would inherit a provider unable to serve the model. The nine ulw-* workflows (ulw-interview, ulw-research, ulw-plan, ulw-work, ulw-maestro, ulw-loop, ulw-qa, ulw-perf, ulw-context) are triggered from chat and routed by Hermes. ulw-work splits an accepted plan into units that never share a file, each in its own worktree, returning typed four-state results. ulw-maestro delegates tasks to Claude Code or Codex CLI, each lane visible as its own HUD row with model, effort, turn, tokens, and cost. Model chains live in ~/.omh/routing/model-chains. and are editable via omh model-chains set quick "kimi-k3-ultrafast:low, glm-5.2-ultrafast:low"; the Maestro lane is configured with omh coding category-maestro set/interview. Long-term memory captures candidates from sessions onto review cards, records only approved items with provenance and a review-due date, and hands the next session a recall pack cut to a token budget. It also provides codebase-uml/omh codegraph uml architecture diagrams, refactor-plan phased refactors behavior-locked by tests, an ast-grep structural search playbook, and user-written toolcall guardrail rules.
- Developers already using Hermes Agent who want stronger routing and workflows without switching agents
- Coding teams that need to split long tasks into non-conflicting parallel lanes and require verification evidence before declaring completion
- Users holding Claude Code subscriptions, Codex, or multiple model APIs who want each category to automatically pick the best-served model
- Long-running project maintainers who want decisions and context to persist across sessions with every memory write reviewed
- Engineers who need a repo-wide architecture diagram and a phased, test-locked refactor plan landing one PR per phase
- Users who want to explicitly hand coding tasks to Claude Code or Codex from inside Hermes via the Maestro lane
What are this agent's strengths and limitations?
- Complexity-scored routing with chain fallback; README-cited same-task measurements show GPT-6 Astra producing the same answers for $0.66 instead of $4.29, in 5 minutes instead of 23
- Prompt calibration per model family (13 families) is measured where a route exists — a first-draft Astra block that cost 10% more for the same answers was cut
- Evidence states are explicit: reported done is displayed separately from Test · verified, so no unverified claim reads as complete
- 108 omh-* specialist skills load into the run as tool calls without ever manually invoking an expert
- Long-term memory is file-backed, review-gated, provenance-tracked, and never reads or patches Hermes' own memory
- All model chains are user-editable in one JSON file or via CLI commands without touching code
- Core operation depends on NousResearch's Hermes Agent as the host; without Hermes the operating layer is unusable
- Model names (GPT-6 Astra, Claude Fable 5.1, Kimi K3, etc.) and cost figures come only from the project's own README with no independent verification
- Advanced features (Maestro lanes, fanout, calibration) create a substantial configuration surface across model-chains., providers., category-maestro., and dispatch-models.
- The README states recommendation chains are editable preferences, not benchmark results; real outcomes depend on which providers and subscriptions you hold
- The install path is remote curl | sh execution, and the agent install protocol requires pinning a commit SHA with step-by-step approval — a heavier setup than a typical plugin
How do you install or deploy this agent?
macOS/Linux: curl -fsSL https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.sh | sh. Windows PowerShell 5.1+: irm https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.ps1 | iex. Also available via Homebrew brew install rlaope/tap/omh, Bun bun install -g oh-my-hermes, or npm npm install -g oh-my-hermes. After any install, run omh setup (required) to complete interactive model and provider configuration; verify with omh doctor. Alternatively paste the repository's English install prompt into your AI agent, which resolves main to a full commit SHA via git ls-remote and executes only the pinned INSTALL_FOR_AGENTS.md.
How do you use this agent?
Run omh to open the Hermes terminal wearing the OMH identity (same door as hermes). Speak triggers in chat to route workflows: have Hermes clarify requirements one question at a time (ulw-interview), do sourced research (ulw-research), build a reviewed plan (ulw-plan), execute in parallel lanes (ulw-work), or loop until the goal passes (ulw-loop). Inspect effective chains with omh model-chains show and change them with omh model-chains set; dispatch via the Maestro lane with omh coding run --owner codex --category ultrabrain --goal .... Ask Hermes to "set up my models" anytime. Override defaults by editing ~/.omh/routing/model-chains.; OMH stores only provider IDs, never credentials.
How does this agent compare with similar options?
The README explicitly names Claude Code and Codex: rather than hiding or replacing coding executors, OMH routes to them as an opt-in second lane through ulw-maestro (with readiness probes, capability snapshots, per-run model/effort), and stresses this is never the default path. OMH itself is an operating layer for Hermes Agent — complementary to these CLIs, not a like-for-like alternative.
FAQ
Do I need Hermes Agent installed first?
omh opens the Hermes terminal. Without Hermes it does not run as an agent.Where do costs come from?
unknown, never $0. OMH itself is free and MIT-licensed.Does it modify my existing Hermes config or memory?
What happens if a provider rejects a model?
What is the difference between an executor saying done and actually done?
Code · reported done means the executor said it finished and nobody checked. Test · verified means a test, review, or CI gate actually passed. An exit 0 with no evidence stays reported done until a gate checks it.