What is oh-my-hermes?
oh-my-hermes (OMH) is an add-on for Nous Research's Hermes Agent, not an agent of its own. It is a community project by the GitHub user rlaope, not a Nous release; the README thanks Nous for Hermes and lists OpenGateway, a model gateway, as its sponsor. OMH installs a pack of Hermes skills, a Hermes plugin and its own terminal interface, which runs on Hermes' gateway. The docs say it never patches Hermes, and Hermes still makes every model call, runs every tool and decides every approval.
"Wired into IDE" in the tagline does not mean an editor extension. The docs describe no VS Code or JetBrains extension and no ACP adapter. What exists is closer to IDE plumbing: optional language-server tools for definitions, references and renames (off by default), an MCP server whose config omh setup can write for Claude Code, Codex, OpenCode or Cursor, and a ulw-maestro workflow that hands a task to Claude Code or Codex and follows the live session.
OMH works with whatever providers Hermes can reach; its site lists 26 model families and providers, from OpenAI and Claude to Bedrock and OpenRouter. Each piece of work is tagged with one of 18 categories, each with an ordered chain of models and a reasoning effort, and a rejected model falls through to the next in the chain. The repository was created in June 2026, is MIT-licensed and mostly Python, and has shipped 22 releases since v1.0.0 on 9 June; v4.0.3 came out on 28 October 2026.
Is oh-my-hermes worth it?
oh-my-hermes is the most ambitious attempt to turn Hermes Agent into a coding harness, and its best idea is a reporting rule rather than a feature. About 4,970 GitHub stars and 400 forks in five months is real interest, though small beside Hermes' roughly 257,000 stars, and npm counted 741 downloads of the package in the week to 29 October.
The core workflow is /dag-ulw. It turns a goal into a graph of small nodes, runs the ready ones in parallel on the model chosen for each node's category, keeps nodes that run at the same time out of each other's files, and has a separate node check the goal before the report. Every status then names its stage and how sure OMH is: Plan · not run, Code · reported done, Test · verified. A run counts as verified only when a test, review or CI gate actually passed. Most harnesses print "complete" when the model says so; this one does not.
On safety, OMH adds little of its own because Hermes keeps control. The docs say Hermes owns every tool run and approval decision, so isolation is whichever Hermes terminal backend you chose, and Hermes' local backend has none. OMH's terminal interface adds an accept-all mode on Shift+Tab that switches off Hermes' approval prompts for the session. Its own guards are narrower: child processes it spawns for parallel work get a least-privilege environment by default, and user-written tool-call rules can block a matching call, though the hook fails open, so an invalid rule is skipped without a warning unless you run the validator. Core omh makes no network calls of its own; the startup update check is opt-in, and its "cost telemetry" is local token and dollar counts.
Against Claude Code or OpenCode it is a different kind of tool. Those are coding agents; OMH is an orchestration layer that, for heavy coding, prefers to hand work to Codex, Claude Code or other oh-my runtimes and record what came back. Its own parity matrix measures it against the oh-my family's common patterns: skill packs, specialist roles, worker teams, worktrees and a status display. The costs are churn and context. Version 4.0.0 arrived on 19 October, four months after 1.0.0; the default full install adds every skill's index line to every Hermes request; and the A/B lane comparing Hermes alone with Hermes plus OMH has a corpus but, per the README, no measured run yet.
Where it earns its place
A developer already running Hermes Agent with keys for several providers
Split a large refactor across models without paying frontier prices for every step
Run /dag-ulw so planning lands on a deep-reasoning category and small edits on a quick one, with chains you reorder through omh model. OMH is free; the bill is the tokens each lane uses, which the status display prices per row or marks unknown.
An engineering lead at a 15-person agency that reports progress to clients
Stop agent runs being reported as finished when nothing was tested
Use the stage-and-confidence labels and ulw-loop with a named gate such as npm test, so a run reads Test · verified only after the gate passes. The risk is setup weight: the full skill catalogue adds context to every Hermes request, so a core install may suit a narrow team.
A staff engineer who uses Claude Code or Codex for implementation
Keep planning and research in Hermes and send the coding to an existing agent
ulw-maestro composes a prompt for Claude Code or Codex and follows the session, and omh setup --with-mcp can register OMH's MCP server with Claude Code, Codex, OpenCode or Cursor. Each agent's own subscription or API costs apply.
Before you commit
- Setup
- curl install script for macOS and Linux, a PowerShell installer for Windows, Homebrew, npm or Bun; needs Python 3.11+ (and Node 18+ for npm or Bun) plus a working Hermes Agent. Run omh setup, then omh doctor to check the install.
- Integrations
- Hermes Agent skills, plugin and gateway; an optional MCP stdio server with config recipes for Claude Code, Codex, OpenCode and Cursor; optional language-server tools; hand-offs to Claude Code and Codex.
- Data and privacy
- Runtime records stay local under ~/.omh and, per the docs, hold metadata rather than raw prompts by default. Core omh makes no network calls; prompts go to the providers Hermes uses. Private-file permissions are not enforced on Windows NTFS. MIT licence.
- Cost at scale
- No paid tier; spend is model tokens across parallel lanes, which multiply with the number of nodes in a run. Per-row dollar estimates come from a rate table in the project.
- Limits
- No editor extension; Windows npm and Bun launchers are not yet covered by CI; four major versions in five months; the default full install raises context cost on every request; the promised Hermes vs Hermes plus OMH benchmark is unpublished.
What’s good
- Status labels separate planned, reported and verified work
- Per-lane model chains that fall through instead of silently downgrading
- Works with any provider Hermes supports
- Core omh makes no network calls; update check is opt-in
What’s not
- No editor extension despite the IDE tagline
- Hermes alone vs Hermes plus OMH benchmark not yet run
- Four major versions since June 2026
- Accept-all mode turns off Hermes approvals for a session; tool-call rules fail open
Best for
Hermes Agent users who want planned, parallel, multi-model coding runs with clear evidence of what actually passed.
oh-my-hermes alternatives
oh-my-hermes: common questions
- Is oh-my-hermes worth it?
- oh-my-hermes gives Hermes the planning, parallel lanes and honest status reports of a serious coding harness, but it is five months old, on its fourth major version and still has no published benchmark showing it helps. We rate it 3.7 out of 5.
- How much does oh-my-hermes cost?
- Free under the MIT licence; you pay the model providers Hermes calls. No paid edition is published.
- What are the downsides of oh-my-hermes?
- No editor extension despite the IDE tagline; hermes alone vs Hermes plus OMH benchmark not yet run; four major versions since June 2026; accept-all mode turns off Hermes approvals for a session; tool-call rules fail open.
- Who is oh-my-hermes best for?
- Hermes Agent users who want planned, parallel, multi-model coding runs with clear evidence of what actually passed.
- How much does oh-my-hermes cost for a team or in production?
- No paid tier; spend is model tokens across parallel lanes, which multiply with the number of nodes in a run. Per-row dollar estimates come from a rate table in the project.
- Is oh-my-hermes safe for company data?
- Runtime records stay local under ~/.omh and, per the docs, hold metadata rather than raw prompts by default. Core omh makes no network calls; prompts go to the providers Hermes uses. Private-file permissions are not enforced on Windows NTFS. MIT licence.
- What are the best oh-my-hermes alternatives?
- The closest alternatives we have reviewed are Hermes Agent, OpenCode and OpenAI Codex CLI.
Answers checked Oct 31, 2026.