Who uses Hermes Agent, and what they give it

Hermes Agent is Nous Research’s open-source agent. People run it on a VPS, a Mac Mini or a Raspberry Pi, and talk to it in the terminal or over Telegram, Discord, Slack and two dozen other apps. It sends more tokens through OpenRouter than any other app, Claude Code included.

Nous publishes no user survey, so we read what is public: the configs users paste into the project’s GitHub issues, OpenRouter’s page for the app, the personality files people commit to GitHub, the skills they install, Nous’s own page of 371 user stories, and about 25 posts by people describing their setup. Together they show who the users are, which tools and models they connect, how they shape its personality and what they hand it to do.

This follows the AI agent landscape and which MCP servers each community uses.

Nous keeps a page of 371 user stories gathered from X, Reddit, GitHub, Hacker News and blogs, each tagged by Nous. Rolled into personas, a third are developers and builders, a fifth run a business with it and a fifth use it as a personal assistant they text.

The page leaves one group out. Among the first-person posts we read, about three in ten are from people running a homelab, who use Hermes as a systems administrator. Nous has no tag for them.

The project’s GitHub issues are the largest public record of real setups: 35,092 issues from 16,843 people since February, many with the user’s own config pasted in. We counted people, not issues, so one prolific reporter counts once.

Chat apps. Telegram is named by more issue authors than any other app, almost twice Discord. Then come Slack and, a surprise for an English-language project, Feishu, WeChat and QQ: Hermes has a large Chinese following. But when the bug form asks which messaging platform the problem happened on, the most common answer is none: 829 people use it only in the terminal, more than chose Telegram, Discord, Slack and WhatsApp together.

MCP servers. Fewer people paste MCP configs than chat settings (1,853 issue authors mention MCP at all), so this ranking is directional. The most-named server is Hermes itself: people serve their Hermes to other agents and editors. Then come the developer basics (GitHub, Context7, a browser, the file system), memory servers bolted on beside Hermes’s own memory, and the personal ones: Google Workspace, Home Assistant and desktop control.

Nous publishes a catalogue of 65 approved servers, all off by default, from Linear, Notion and Stripe to Figma and Sentry. Users mostly bring their own: Home Assistant, desktop control and the memory servers are not in it.

What they work on. Memory and scheduled jobs are what people write about most, ahead of skills: the agent that remembers and works on a timer is the one they are building. Most run it on Linux, and a quarter of those who say are on Windows.

Hermes works with any model, and its users pick cheap, open ones. Of the 53.56T tokens Hermes users sent through OpenRouter in the last 30 days, DeepSeek’s models took 34%. Claude Opus, the top model until March, is now eleventh with 1.8%. Nous’s own Hermes 4 models hardly appear.

OpenRouter sees only part of the traffic. Users who sign in with a ChatGPT or Claude subscription, use Nous’s own portal or run a local model never pass through it, and the posts we read show many doing exactly that: GPT-5.5 on a ChatGPT plan, GLM through Z.ai, Qwen and Gemma on their own GPUs.

The GitHub issues show the traffic OpenRouter misses. Claude is the model family issue authors name most, 1,800 of them, ahead of DeepSeek. In the configs people paste, the most common providers are their own OpenAI-compatible endpoint (300 people) and a ChatGPT Codex sign-in (126), both ahead of OpenRouter (108). Nous’s own Hermes models are named by 24 people.

A skill is a written procedure the agent loads when a job calls for it, and Hermes also writes its own. Nous ships 211: 58 bundled, most of them for productivity and software work, and 153 optional ones, led by machine learning ops (36), creative work (31) and research (15). Hermes can install more from skills.sh, ClawHub, LobeHub and GitHub.

Installs are small next to the token counts. Nous’s most-installed skill on skills.sh, dogfood, which has the agent test the software it builds, has 6,300 installs; on ClawHub the most-downloaded skill of any agent has 484,868. Most users let Hermes write skills from what it does for them, and several complain it writes too many and overwrites their edits.

Hermes takes its character from a file called SOUL.md, the first thing in every prompt. A /personality command adds presets for a session, from concise and technical to pirate and noir. On GitHub, 2,200 SOUL.md files mention Hermes, in at least 763 repositories, and 285 of them sit in a profiles/ folder: one person running several agents, each with its own character.

We read 69 of them, picked at random. A third are Nous’s default. The rest are written: named assistants, roles in a company of agents, and operators for one business. They share a style. Most ask for directness, no filler and no emoji, and many write their limits into the character itself: read-only, never place a trade, ask before acting. One in eight is not in English.

Scheduled jobs are the feature users mention most. The typical setup is an agent that works while its owner sleeps and reports by message in the morning: a briefing, a ranked to-do list, the state of the servers. Most of these jobs read and draft, and leave sending, paying and fixing to a person.

Nous chose the stories. Its page is curated, so it leans toward successes and builders, and the tags are Nous’s.

Reddit, X and Discord are mostly out of reach. Reddit and X blocked our readers and Discord is not public, so they appear here only where Nous’s page or a search result quotes them. People who write blog posts are more technical than most users.

GitHub issues are written by people with a problem. A config pasted into an issue shows what someone connected, not what works or what is used most. Many issues read as drafted by an agent, and counts that match a word (memory, Claude) also catch code and comparisons; the bug form’s own boxes and the project’s labels are cleaner, and we show those where they exist. We read issue titles and bodies, not comments.

hedwigai publishes skills that work in Hermes: web traffic and Google rankings, lead lists and hiring signals, SEC and FDA records, and a reference checker. Each answers from public records, with a source cited for every claim. Hermes can also reach hedwigai as an MCP server.

npx skills add un-seen/hedwigai-skills