Hermes Agent takes a different shape from the other coding agents, and that changes how you get recommended by it. Built by Nous Research as an open-source, model-agnostic life-OS agent, it runs on whatever model the user configures and extends through an installable skills system. So two levers matter at once: the general training-data signal that follows you across whatever model someone picks, and a SKILL.md package that makes your tool a first-class capability Hermes can install. This is the playbook for getting recommended by Hermes Agent, sitting under the broader pillar on generative engine optimization, extended from citation to agent tool-selection.
Key takeaways
- The distinctive lever — The skills system. A tool packaged as a SKILL.md and submitted to the community registry becomes a first-class installable capability — the direct distribution channel.
- The inherited lever — Model-agnosticism means many recommendations come from whichever model the user configured, so general training-data prevalence and good docs still carry weight.
- The retrieval lever — Hermes pulls the live web via Firecrawl and supports MCP, so a clean README, an llms.txt, and an MCP server all improve what it finds.
- The positioning — Hermes is a persistent life-OS agent more than a coding one, and often orchestrates other agents for code — so target it as an orchestration and workflow surface.
How Hermes Agent chooses what to recommend
Hermes blends an inherited model layer with its own extensions and retrieval. The model-agnosticism is the part that makes it distinct.
- Inherit — the user-configured model brings its own training-data priors to any recommendation.
- Install — skills packaged as SKILL.md become first-class capabilities the agent can use.
- Retrieve — the Firecrawl integration pulls the live web; MCP servers are callable too.
- Recall — a hybrid retrieval layer draws on the user's own documents and context.
- Recommend or orchestrate — Hermes suggests a tool, or hands a coding task to another agent.
Two implications follow. First, because you cannot pin down which model a Hermes user runs, the durable baseline is broad training-data prevalence rather than tuning for one provider — be well-represented everywhere, not optimised for one cutoff. Second, the SKILL.md skills system is the lever Hermes actually hands you: a published, installable skill is the difference between being a tool Hermes might mention and one it can directly use, which is the strongest form of recommendation on any agent.
The playbook
Tactics ordered by leverage, calibrated for Hermes. The first is Hermes-specific; the rest are the cross-agent fundamentals applied here.
- Package your tool as a SKILL.md and submit it to the registry. This is the direct distribution channel. A SKILL.md in the community-registry-compatible format makes your tool an installable Hermes capability, and submitting it to the skills registry puts it where Hermes users discover and add capabilities. It is the equivalent of Claude Code's MCP server for this agent.
- Maintain broad training-data prevalence. Because Hermes is model-agnostic, you want strong GitHub, npm, and PyPI signal and wide documentation that follows you across whatever model the user runs. This is the inherited-layer lever, and model-agnosticism makes breadth matter more than depth on any single provider.
- Keep a clean README and llms.txt for Firecrawl retrieval. Hermes pulls the live web via Firecrawl, so well-structured docs and an llms.txt improve the accuracy of what it retrieves about your tool — the same assets that serve the text engines and the other agents.
- Reuse your MCP server. Hermes supports MCP, so an MCP server built for Claude Code is usable here too. Where you have one, it complements the SKILL.md rather than competing with it, giving Hermes both an installable skill and a callable server.
- Target the orchestration and workflow context. Hermes is a life-OS and orchestration agent more than a pure coder, often handing coding to other agents. Position your tool for the workflow and automation tasks Hermes itself runs, not only the code tasks it delegates, so you are relevant to what Hermes does directly.
- Describe capabilities in conventional, legible terms. A SKILL.md whose description clearly states what the tool does, in the vocabulary users and models already use, is easier for Hermes to select and apply than one with an opaque name or vague scope.
What's different from Claude Code, Codex, and OpenClaw
The packaging format and the model model differ across the four agents; the baseline of strong public signal and clean docs is shared.
- Claude Code centres on MCP servers and Context7; Hermes uses SKILL.md skills but also supports MCP, so there is overlap. The Claude Code playbook is at get recommended by Claude Code.
- Codex centres on AGENTS.md; Hermes does not, so the levers diverge. The Codex playbook is at get recommended by Codex.
- OpenClaw shares the SKILL.md model through its ClawHub registry, so a skill built for one is close to a skill built for the other. The OpenClaw playbook is at get recommended by OpenClaw.
- Model-agnosticism is Hermes's own wrinkle: with no fixed model, breadth of training-data presence matters more than optimising for any single provider's cutoff.
Measurement
Skill installs are the cleanest signal Hermes offers. Build the loop in three layers:
- Skill installs and usage. If you publish a Hermes Skill, track installs and usage from the community registry — the most direct evidence your distribution channel is working.
- Direct testing. Run Hermes on representative tasks and see whether it reaches for your tool, re-testing after publishing a skill or improving docs.
- Cross-reference the text engines. The docs-and-llms.txt work that helps Hermes also helps the answer engines, so an LLM-visibility tracker gives a related read. The Radar's shortlist is at 6 GEO Tools the Radar Actually Recommends; CTAIO Labs tested ten in the visibility tools test.
Related reads
Frequently asked questions
What is Hermes Agent?
Hermes Agent is an open-source, MIT-licensed agent from Nous Research. It is model-agnostic, routing through whatever model the user configures — via OpenRouter's many models, NovitaAI, NVIDIA NIM, or any OpenAI-compatible endpoint — and it is positioned more as a persistent personal life-OS agent than a dedicated coding tool. For coding tasks specifically, it often orchestrates other agents. It retrieves the live web through a Firecrawl integration, supports MCP, and extends through a skills system.
How does Hermes Agent decide what to recommend?
Partly through the underlying model and partly through its own extensions. Because Hermes is model-agnostic, many recommendations are inherited from whatever model the user selected, so that model's training-data priors apply. On top of that, Hermes can install skills, retrieve the live web via Firecrawl, and call MCP servers. There is no published ranking spec, so this is a description of the inputs rather than a disclosed algorithm — treat the specifics as inference to test.
What is the SKILL.md skills system and why does it matter?
It is Hermes's first-class extension format. A tool packaged as a SKILL.md in the community-registry-compatible format becomes an installable capability the agent can use directly, rather than something it merely knows about. Submitting your tool as a Hermes Skill to the community registry is the direct distribution channel for this agent — the equivalent of an MCP server for Claude Code or an AGENTS.md for Codex, and the highest-signal lever you have here.
Does model-agnosticism change my strategy?
Yes, in one important way: you cannot optimise for a single fixed model, because the user chooses it. That means general training-data prevalence — strong GitHub and package-registry signal, wide documentation — matters across whatever model they pick, rather than tuning for one provider's cutoff. Combine that broad baseline with a Hermes Skill for the agent-native layer, and you cover both the inherited and the Hermes-specific paths.
Does an llms.txt or MCP server help with Hermes?
Both feed it. Hermes retrieves the live web through Firecrawl, so a clean README and an llms.txt improve the accuracy of what it finds, and its MCP support means an MCP server is usable here too. The practical upshot is that the agent-native assets you build for Claude Code — an MCP server, good docs, an llms.txt — are not wasted on Hermes; they complement the SKILL.md that is its distinctive channel.
How is getting recommended by Hermes different from Claude Code or Codex?
The packaging format and the model situation differ. Claude Code centres on MCP and Context7, Codex on AGENTS.md, and Hermes on SKILL.md skills — three first-class formats for three agents. And Hermes's model-agnosticism means it has no single training cutoff to plan around, unlike Claude Code or Codex which run fixed model families. The shared baseline — strong public signal and clean docs — still helps across all of them.
How do I measure whether this is working?
Through skill installs and direct testing. If you publish a Hermes Skill, track its installs and usage from the community registry as the most direct signal. Test by running Hermes on representative tasks to see whether it reaches for your tool, and correlate broader installs and sign-ups with agent-driven discovery. As with the other agents, there is no recommendation dashboard, so measurement is proxies plus direct tests.
What is Hermes Agent?
Hermes Agent is an open-source, MIT-licensed agent from Nous Research. It is model-agnostic, routing through whatever model the user configures — via OpenRouter's many models, NovitaAI, NVIDIA NIM, or any OpenAI-compatible endpoint — and it is positioned more as a persistent personal life-OS agent than a dedicated coding tool. For coding tasks specifically, it often orchestrates other agents. It retrieves the live web through a Firecrawl integration, supports MCP, and extends through a skills system.
How does Hermes Agent decide what to recommend?
Partly through the underlying model and partly through its own extensions. Because Hermes is model-agnostic, many recommendations are inherited from whatever model the user selected, so that model's training-data priors apply. On top of that, Hermes can install skills, retrieve the live web via Firecrawl, and call MCP servers. There is no published ranking spec, so this is a description of the inputs rather than a disclosed algorithm — treat the specifics as inference to test.
What is the SKILL.md skills system and why does it matter?
It is Hermes's first-class extension format. A tool packaged as a SKILL.md in the community-registry-compatible format becomes an installable capability the agent can use directly, rather than something it merely knows about. Submitting your tool as a Hermes Skill to the community registry is the direct distribution channel for this agent — the equivalent of an MCP server for Claude Code or an AGENTS.md for Codex, and the highest-signal lever you have here.
Does model-agnosticism change my strategy?
Yes, in one important way: you cannot optimise for a single fixed model, because the user chooses it. That means general training-data prevalence — strong GitHub and package-registry signal, wide documentation — matters across whatever model they pick, rather than tuning for one provider's cutoff. Combine that broad baseline with a Hermes Skill for the agent-native layer, and you cover both the inherited and the Hermes-specific paths.
Does an llms.txt or MCP server help with Hermes?
Both feed it. Hermes retrieves the live web through Firecrawl, so a clean README and an llms.txt improve the accuracy of what it finds, and its MCP support means an MCP server is usable here too. The practical upshot is that the agent-native assets you build for Claude Code — an MCP server, good docs, an llms.txt — are not wasted on Hermes; they complement the SKILL.md that is its distinctive channel.
How is getting recommended by Hermes different from Claude Code or Codex?
The packaging format and the model situation differ. Claude Code centres on MCP and Context7, Codex on AGENTS.md, and Hermes on SKILL.md skills — three first-class formats for three agents. And Hermes's model-agnosticism means it has no single training cutoff to plan around, unlike Claude Code or Codex which run fixed model families. The shared baseline — strong public signal and clean docs — still helps across all of them.
How do I measure whether this is working?
Through skill installs and direct testing. If you publish a Hermes Skill, track its installs and usage from the community registry as the most direct signal. Test by running Hermes on representative tasks to see whether it reaches for your tool, and correlate broader installs and sign-ups with agent-driven discovery. As with the other agents, there is no recommendation dashboard, so measurement is proxies plus direct tests.
Ready to Find the Right AI Tools?
Browse our data-driven rankings to find the best AI tools for your team.