Cerita asli dari X, GitHub, Reddit, Hacker News, YouTube & blog — dikelompokkan per kategori fitur. Sumber: halaman resmi User Stories.
I built Hermes Swarm, a self-hosted server that runs teams of AI agents 24/7. Each agent is a full Hermes agent. That means they have their own terminal, a browser and read/write access to the team's shared filesystem. They message each other peer-to-peer, delegate work and self-schedule their own wake-ups.
I built Hound as a drop-in MCP server that handles all of it. No API keys, no Docker, no config wizard. Just add it to your `~/.hermes/config.yaml` and your agent gets 6 web research tools.
The version that works is much simpler: 1. I create saved searches on sites like Zillow and Trulia. 2. New-listing alerts go to a dedicated Gmail account used by my Hermes agent. 3. Once a day, a cron job reads the alerts through the Himalaya email CLI. … automate the 97% that involves gathering and filtering information
Hermes IdleViber is an ambient idle/incremental game that lives in your browser. … The twist? The game is hard-locked to your local Hermes Agent gateway.
My agents were hitting 403s constantly. Cloudflare, Akamai, Datadome — if the site has any bot detection at all, a datacenter IP and a headless browser get blocked immediately. … Five Hermes Agent instances on my homelab, all routing through a residential IP via a Raspberry Pi on a home network.
What started as a research task turned into full-blown community project It spent the week mapping every tool, skill, and integration built for @NousResearch's Hermes Agent
My Hermes agent by @NousResearch runs on an old @DellTech @DellXPS laptop with a busted GPU. I bought it 5 years ago just for Zoom calls and Spotify. It still works.
Hermes Agent: Everything else. Blog posts, competitor research, news monitoring, code reviews. The grunt work that needs speed.
Connect your Nous Research Hermes Agent to Hindsight Cloud, the best-performing AI Agent memory, in a few easy steps!
told it to Google me and then build a landing page based on what it found and that was genuinely mind blowing because it ran the searches, found kinks, created the page, SSH'd into my VPS, uploaded the page, then texted me when it was done. what?!
I set up my Hermes Agents with local SearXNG (in a container that they share) before ddgs (duckduckgo) was added, because I thought all API use of DDG was paid. I've been pretty happy with searxng so far.
Full copy-paste setup for Hermes Agent + Browser Harness on a Hostinger VPS. Register Browser Harness as a Hermes skill via symlink so Hermes can find and use it. Recommended model: anthropic/claude-opus-4.7 via OpenRouter.
'Research the top trending AI tools right now and come back with the top three that would make for an interesting tutorial video. Create a new skill based on your approach and call it YouTube-video-research. Can you set up a weekly job that runs every Monday at 9:00 AM using that skill?'
I got tired of paying Perplexity api to use their mcp for agentic research. It was like $10 a pop every few days just for synthesis calls on the research dumps of my other mcp tool calls. So I made my own called Gigaxity. Gigaxity is comprised of 7 mcp's and some companions like SearXNG (with custom config yaml for optimal websearch settings). The context sources are picked in such a way as to maximize free tier api keys.
Claude (Opus 4.7) handles my day-to-day chats: research, planning, working through problems. Hermes runs 24/7 on a mini PC and handles the real-world stuff: email, web browsing, form filling, calendar updates, cron jobs.
Higgsfield Marketing Studio powered by Hermes Agent is doing the replacing this time. Paste product URL → Hermes scrapes the landing page, pulls winning ad hooks from Meta Ads Library + TikTok Creative Center in the exact niche, and writes the brief itself. Total time: ~4 minutes.
I mapped all open + closed PRs to find what was truly missing. PR #64 added MCP client support (Hermes → external tools). The server side was completely absent. So I built hermes mcp-server — making Hermes a full MCP server so Claude Desktop, Cursor, and any MCP client can use Hermes's tools directly. Exposes 9 Hermes tools over MCP: terminal, read/write file, web search, memory, skills, run\_agent.
Sometimes Hermes Agent melts my heart @NousResearch.
This research is gold. Been deep in Hermes for weeks and it's started autonomously suggesting entire workflows I never would have designed myself. One it built for me recently: X API to extract from lists and bookmarks to structure into article to NotebookLM podcast. I'm building a physical AI companion with Hermes as the core cognitive layer right now.
There's one Hermes use case for everyone — build a research agent. Mine watches the AI/agent space, picks out useful signals, writes briefs, suggests content angles, tracks what I ignore, and keeps improving its own workflow. Delivers daily via Discord, Slack, Notion, email, Obsidian, and local markdown.
hermes-lab is the bookkeeper for running experiments autonomously. you give it a search space and a way to evaluate results, it handles scheduling, tracking, and suggesting what to try next. inspired by the autoresearch wave (karpathy, sakana ai scientist, AIDE) but generic — works for ml training.
I built a dashboard that makes Hermes config and management much easier. Browser-based terminal with real PTY, file explorer + inline editor, agent management: switch profile, start/stop gateway in one click, realtime log streaming, token usage analytics. Designed for single-user local networks or reverse-proxied VPS deployments.
Apple Health: Hermes wrote Python on the fly and found my sleep avg was 7.59 hrs. Threads Analytics: drop cookies in, pulled 34 posts of analytics in one command. Hermes is dramatically better than OpenClaw at browser automation. Gmail + Calendar OAuth via drag-drop JSON. Hermes = CEO, OpenClaw = Senior Engineer, both pointed at the same Obsidian vault on my NAS.
I'm using Hermes. The same applies to all agents, don't give it free reign over all your stuff. Run it within a sandbox. https://github.com/nousresearch/hermes-agent
I've switched to Nous-Research Hermes-Agent from previous Agents I've been experimenting with. Hermes is unique in that it self-learns. Within a single day, I built and launched five small applications.
Hermes Agent by Nous Research can now scrape, search, and interact with the web using Firecrawl. Enable it during setup to give Hermes the ability.
memory tool — Save/retrieve durable facts across sessions. I save user preferences, environment details, tool quirks, stable conventions. It's injected into every turn. I keep it compact — facts that will still matter later. I do NOT save task progress or temporary state here. session\_search — Recall past conversations. skill\_manage — Save procedural memory. After complex tasks (5+ tool calls), tricky fixes, or non-trivial workflows, I save the approach as a skill.
Conversation continues across platforms (Telegram, Discord, Slack, WhatsApp, Signal, terminal). Real memory: two curated files MEMORY.md + USER.md, plus SQLite full-text search over all past sessions. Scheduled tasks via natural language — no crontab editing.
ChatGPT is a browser tab. Hermes is a server process that's running right now, has been building a model of your workflow for the past few weeks, and just sent you a Telegram message before you woke up. It doesn't stop when you close your laptop. It messages you first. Memory that gets useful over time. Runs code, doesn't just write it. Takes action in your actual apps.
I created a beautiful web interface for Hermes Agent that adds dark/light theme, persistent memory using MEMORY.md and USER.md, per-session chat history, status bar, responsive on mobile and desktop.
Actually I've been building exactly this for my agent (Hermes) for a while now... Instead of a dumb vector search, my memory kernel intercepts the LLM's thoughts and 'compiles' them into a living structured database. Every extracted memory gets a lifecycle status. If the agent learns something that contradicts old data, a background orchestrator automatically detects it.
Using both side by side I find it more stable and less headache than OC. Hermes has more research skills, and it's very handy as troubleshooter of OC. Telegram recently added bot to bot communication in their API so I'm thinking a way both communicate.
Hermes Desktop is a native Mac app for Hermes Agent. Not a browser wrapper but a real SwiftUI app that connects to your Hermes host and files over SSH. The gateway model is the right answer for Telegram and Discord. For a native Mac app that sits next to your terminal, it's just not the one I'd ship.
It set up cron jobs that search for news/leaks/rumors in the tech space, then created channels on Discord by importance/urgency. It auto-contextualizes each news item to my vault and the actual work I have across video projects — I get up-to-date insights and tweak videos to stay super relevant. Updates 3x a day, always learning and adapting.
Users want secure remote access to the Hermes API server / Open WebUI without exposing ports publicly. Tailscale serve provides zero-config HTTPS tunneling over a private mesh.
Built a memory provider plugin connecting agentmemory to Hermes. Covers cross-agent memory (developer using Hermes plus Claude Code or Cursor) with hybrid BM25+vector+knowledge-graph search.
Chat via Telegram while execution runs on Modal serverless (cheap when idle). Run on a $5 VPS that stays up when the laptop closes. Pin to SSH backend inside a customer's VPC for consulting. Verified benchmark (TokenMix): self-created skills cut research-task time by ~40% vs. a fresh agent.
The NousResearch team built Hermes Agent — an open-source agentic AI system with tools, skills, memory, and multi-platform messaging. It's good. So I built a native Windows app around it.
Being an avid music lover and using Tidal I've created this skill set up: tidal-curator (Base Skill) for OAuth auth, playlist management, and track search; tidal-dual-curator for automated twice-daily music curation at 9am (morning vibe) and 6pm (evening vibe), with time-of-day matching, artist background, and no duplicates.
I recently participated in Nous Research's Hermes Agent Hackathon, which is where this project was born. Every tool out there assumes a human operator with automation bolted on. I wanted to flip that — make the agent the primary driver and give the human a supervisory role.
I was an early user of Hermes Agent and have been a power user ever since. Scraped every GitHub repo related to Hermes, filtered out unfinished, built an ecosystem map and published a website (hermesatlas.com) where you can see all projects organized by category with star ratings.
Of the 95% of people who use AI, only 5% see real results. Hype's VP of AI, @glitch\_, on why the gap isn't the models, but the architecture. Read his deep dive on building with Hermes Agent, @NousResearch, agent swarms, experiment loops, and what actually compounds.
Vibe after replacing everything with a Hermes agent: autoresearch, Karpathy LLM wiki second brain, skills creation, scheduled jobs, background monitoring, LLM model selection, Telegram/Discord support. A personal automation agent that lives on a server and talks to you through messaging apps or CLI.
I've developed and released HermesAI Translator, an open-source, AI-powered browser extension for seamless translation, reading assistance, text polishing, and summarization directly within the browser, driven by Hermes-4-70B.
With Hermes (built by @NousResearch) providing 40+ built-in tools, persistent memory, and subagent parallelization, the development experience is best-in-class. Built for operations like inventory tracking where context, memory, and real-time inputs are non-negotiable.
Had my (Hermes) research agent dig into what people are actually building with Hermes — turned up an ecosystem mosaic of trading bots, personal assistants, content pipelines and self-hosted everything.
My @NousResearch hermes-agent can make movies now using @browser\_use skill. No API needed. No human intervention. I told it to set the mood, action, camera movement, dialog and overall story — it used Browser-Use and Seedance 2.0 to generate a video.
Client research before calls saves 20–30 min every time. Meeting notes → follow-up drafts. Weekly podcast digest replaced 10+ hrs of listening with a 2hr Hermes workflow using Voxtral. Daily news briefings to Telegram/Discord. Content-ops pipeline (blogs, cold emails, lead scraping from YC, Twitter, Reddit). 24/7 assistant + watchdog.
Successfully scraped Amazon (notoriously difficult) without additional config. Free speech-to-text via local Whisper, free TTS via Edge TTS. YouTube title generator skill produces five search-based, five browse-targeted, and five hybrid titles.
Hermes Agent = automated TikTok slideshow factory. It researches the topic, writes the script, formats the slides, outputs in the exact dimensions TikTok needs. 10 slideshows a week manually takes 30-40 hours. With Hermes it takes the time to review and approve.
Someone posts a coding setup, someone posts an automation, someone posts an article they wrote with it. Same tool, different category, every day. Hermes is not a coding agent. Not a research agent. It is the general agent — one tool running every category of work a builder does.
This guy literally open sourced a full web UI for Hermes Agent. 1:1 terminal parity right in your browser, zero build steps, a sleek 3-panel layout for chat, sessions, and file browsing. 100% free and open-source.
Run three fully isolated agents — a programmer (uses Claude Code), a deep researcher, and a designer — each with its own personality, memory, skills, and Telegram bot. Hermes packages a gateway around a learning agent.
C2 via Telegram: launch tasks from your phone. It autonomously researches SMS gateways, creates a custom SMS skill, deploys a tracking page that harvests GPS, and makes everything publicly accessible using a Cloudflare Quick Tunnel — for ethical red-team engagements.
Monitor Polymarket categories for unusual movement: crypto, elections, AI, sports, macro. Do not place trades. When a market moves more than 8% in 24h or volume spikes, send me a brief with the link, price, possible reason, and related news.
Used Dell OptiPlex from Facebook Marketplace — $175, 16GB RAM, Linux Telegram as the interface (I barely touch the desktop app) Runs 24/7, no Mac Mini, no GPU rig … Every morning at 8:30am I get an email with 1-3 curated rental listings. It filters by school district, commute time, lighting, bedrooms, budget.
Hetzner VPS: €4/month, EU-hosted (or wherever you/your customers are located), everything in Docker containers (isolated, reproducible) \* Hermes Agent: deployed via Ansible playbooks (one command, fully automated) … One of my customers uses it to track ~50 dividend ETFs; the agent queries his tracking sheet, cross-references latest yields, and posts a summary to his Telegram every morning.
Does every single run(setup as scheduled task/job by asking Hermes) cost me LLM tokens? … --script --no-agent : this is the one people(at least beginners like me) miss. The script is the job. No agent loop, no LLM call, period.
I had it pull the top 20 pet poisons, and determine proper treatments (Charcoal vs. Emesis vs. Intralipids vs...) and make a spreadsheet, saved in my Obsidian vault. I can never remember what toxins are treated with what... \*yeah, I verified them all\*
I use CheapCharts daily to lookup deals and pricing for movies and show bundles. I worked with my hermes agent to create a skill using CheapCharts public API.
Every agent commits to a single model, and every model has blind spots the others would have caught. … The unit you work with is a preset. Think of it as a recipe that names a few models to consult and one model to write the final answer, saved under a label you can reuse.
Most people overspend on AI in three places: 1. Defaulting to frontier models for every task
Day 10 of Building AGI for my Hermes Agent: Codex saved the day as a runtime monitor for my agent-to-agent workflows. I used Codex with GPT-5.4 on extra-high to watch the workflow run, catch where it broke, and fix it live until it worked reliably.
I built a monitoring dashboard to profile token consumption on a Hermes v0.6.0 deployment running Telegram + WhatsApp + Cron gateways. After analyzing 6 request dumps, I found that 73% of every API call is fixed overhead.
The jMunch MCP suite provides three MCP servers bringing token-efficient code intelligence (52 tools via tree-sitter), documentation retrieval, and tabular data analysis. Plug-and-play with Hermes's native MCP client.
One of the apps I'm working on helps my friend who owns a remodeling company find work. Specifically roofing work. It started after I saw a video of a guy using Chatgbt and Replit to vibe code some app. I just wondered if I could do it. But the cost of Replit is killing me and the project isn't even a full on CRM lead gen product.
Adding Ombre underneath creates a production-ready stack: tamper-proof audit, prompt-injection blocking, memory encryption at rest, hallucination detection, cost tracking, EU AI Act compliance exports.
I have Hermes running on a Pi4. It saves my preferences while working on tasks like modifying files. I want to use it as a central brain shared across all my devices.
100% human generated. Includes what I use Hermes agent for (since I've seen a lot of people wondering what to do with this thing), and what models/providers I use to keep things cheap. I have been running a multi agent setup for Hermes agent for the last several weeks. It sends me messages on Telegram to remind me.
One of the problems I run into with Hermes is high cost when context windows fill up. One method to fix this is heavy use of sub-agents, but permissions/scope delegation to the sub-agent is often problematic. ZeroID is an agent identity layer that uses RFC 8693 token exchange to handle scope delegation with sub-agents.
Made a network based kanban for agents with local client to minimize token use. Comes with web ui so meat based agents can see what is going on.
your AI agent, but it dreams at night — 5 REM cycles, 23:00–06:00, zero cron jobs. by morning it's written 9 dream thoughts + a recall you can query. costs ~$0.014/night on Haiku. free on local models.
Day 8 of Building AGI for my Hermes Agent: Auto-Build saved me loads of time and tokens. Main agent (GPT-5.4) breaks a plan into phases, coder agent (MiniMax M2.7) implements, QA agent (local Qwen 35B A3B) tests. Plan → implement → test → fail → repair → ship.
Switching to Hermes with OpenRouter cut my token spend ~90% — from ~$130 per 5 days to ~$10 per 5 days. Hermes runs on a cheap Android phone via Termux + Termux API — unlocks SMS, sensors, and on-device social posting. Customization is a trap; output is the skill.
OpenClaw setup: Mac Mini M4 ($599) + Opus 4.6 = ~$80–150/mo. Hermes on VPS: under $20/mo total using Minimax M2.7. Example first task: 'check the top 5 trending GitHub repos right now and send me a summary.'
For executive function, seeing as that's what I'm struggling with, I can't rely on me being the one to prompt it into existence or check over things, because otherwise maybe I wouldn't need the tool in the first place. At the moment my simple cronjobs checking google tasks for changes and sending message to discord and signal are costing 14k tokens.
I'm on day 297 of my streak: 900,000+ seconds of compute time automated, 5,000,000,000+ tokens generated, $100,000+ in client work value automated.
Prompted Hermes to help write tweets based on past video scripts. Pointed it at a scripts folder; it analyzed my writing style, produced usable tweets, and saved preferences to memory automatically. Brand new session test: it recalled everything, including preferred emojis.
Set this as your Smart routing default (using OpenRouter): Tier 1 Hermes (Gemini 3.1 Flash Lite) for clear mechanical multi-file work. Tier 2 Sonnet for ambiguous, delicate, high-risk tasks. Tier 3 Minimax for low-overhead. Run the minimax-cache-optimization skill. Seriously, do this from day one and you'll save about 10 hours of trial and error.
Having a competent agent with constant state has been good for memorializing and organizing important info directly into Obsidian, planning, and working out bugs with my home automation setup. Also helpful dealing with several miscellaneous servers in the house. I have it running on a cheap VPS and it's fairly locked down.
I made this repo to can use Hermes (or opencode) with multiple Agents. But using ollama that is cheaper than other providers.
POV: you use Hermes agent to fix your OpenClaw to save countless hours and credits every day. The setup that saved me hours every day: OpenClaw + Hermes watchdog.
We built an RTK integration for Hermes that transparently rewrites terminal commands to their RTK equivalents, saving 60-90% of tokens in your context window.
Hosting the agent costs nothing. Running the agent the wrong way costs a fortune. Take the default setup at face value and you end up with a working agent and a $400 OpenRouter bill. I rebuilt my personal automation stack on Hermes.
I've been running a 24/7 AI agent on a Mac Mini for 2 months. 18 cron jobs, 35 scripts, 6 custom skills, a structured context system that makes every session smarter than the last. Total cost: $21/month.
A skill is a cache for reasoning, not just data — re-derivation cost drops toward zero. But the same loop compounds cost drift, skill rot, and trust surface. One challenge entrant reported a $47 surprise bill from an overnight run. Skill-cached runs were ~63% cheaper than re-derivation.
In a new Telegram conversation I said 'help me plan a trip to France' and the agent referenced the past Japan conversation, recalling my preferences — nature, history, mid-range budget. That's the difference between a chatbot and a personal assistant.
I want you to schedule a task for yourself to do every single night at 2am. It should be a micro app, UI, or system that helps us get closer to my personal goals — save us time, make us more productive, or just be generally useful.
I run 28 cron jobs and 30+ custom skills. Every single one was built with Hermes, not downloaded. … Every night at 3 AM while I sleep, Hermes reads every single conversation we had. It extracts the decisions made, projects worked on, bugs chased, people talked to, and mistakes that should not be repeated.
Job: Cybersecurity Analyst (DFIR) / Cyber Threat Intelligence Profiles are a game changer. Use them. My default profile is for work. It's connected to my second brain in Obsidian and synced through Git.
I built a custom Document Structuring Skill for Hermes Agent because running massive tech manuals on local hardware was completely blowing up my context window … Test PDF: AMD Technical Documentation \* Size: 3,799 pages
My Hermes agent lives on a VPS. I wanted it to run commands on my work-laptop, read files, deploy code — but my company laptop blocks installs and some URLs. No SSH tunnels, no VPN, no inbound ports. So I built a brain-and-arm architecture
I run four different AI agents day to day: Claude Code, Codex, Grok, and a self-hosted Telegram/Discord/WhatsApp gateway (Hermes). Each had its \*own\* memory. The same facts about my work (people, companies, projects, preferences) lived in 4+ places, drifted apart, and there was no single place to ask. Classic agent amnesia, times four.
You have a great conversation on Monday, and by Friday your agent looks at you like you're a stranger who wandered into its context window. … So I built Synapse a temporal knowledge graph memory system for Hermes agents.
I've been building a local AI workstation over the last few days and wanted to share the final configuration because one optimization completely changed the result.
I've been running my Hermes agent with `Gemma 4 26b a4b q4 qat` on my AMD 7900 XTX gpu, with 15GB size, it leave plenty of room for 200k context length
I connected my entire business to Hermes x Obsidian. Client notes, SOPs, meeting logs, business decisions - all of it. My Hermes agent now runs automations I didn't even know I needed, and my agent self-evolves over time.
I just watched an AI agent destroy a codebase in 3 seconds. … So I typed: /rollback 1 Everything came back. Files restored. Agent context rewound. Like it never happened.
once a workflow is solid, break it out into its own agent with its own credentials, memory and scope.
I've been iterating on a local-first memory/cognition layer for Hermes and finally pushed it to public visibility.
It's right now integrated with SourceDev to index repos (self hosted), Tenderly MCP to debug onchain transactions, LLM-Wiki ingest of Litepaper of our protocol and other docs. Hopefully team will find it useful and will integrate more infra tools overtime to help the team.
I do a really simple journal and log at the end of every day. I tried using Kimi 2.5 just like I would a Sonnet 4.6 but it messed all sorts of things up. When I said 'This is my log for last Thursday. Log it in Obsidian,' it didn't kick off my journaling skill call.
Why Hermes Agent has emerged as the leading open-source AI agent that developers and builders are choosing — self-improving skills, three-layer memory architecture, real-world applications including video dubbing workflows.
Hermes hud is a TUI dashboard that watches your ai agent think. it reads from your agent's memory, tracks skills, sessions, corrections, projects, cron jobs, all of it. live.
Yo! Do you know how I spent my last 200-400 hours? Yeah... I wrote a fking memory kernel for hermes. Why this much of a time? After 3 failed attempts (149 hours of work... for real) I just realized it's a fucking difficult job. So instead of reinvent the wheel I built together the bests for my usecase: 3 layers — L1 Hindsight, L2 Graphiti, L3 MemPalace.
I built a Raspberry Pi 5 to run Hermes on 24/7. it's my first time diving into trying a real AI agent and so far I'm enjoying it. Much better than open claw. The thing is, since Hermes is learning so much about me and my workflows and custom skills, I'd love if I could use Hermes (with its knowledge and memory) on my Mac Studio which I work off of.
I speak Turkish with it... initially said him to 'now i want you to use obsidian to manage my tasks and other stuff. and cross-check with my apple calendar'. later while working on another task today i said 'okay looks good. you know how you will manage tasks and events apple calendars + obsidian + cron/signal' and it confirmed the workflow.
Extending Hermes to full Feishu ecosystem coverage: Documents, Sheets, Bitable, Calendar, Tasks, Wiki, Contacts, Drive, Email. Giving Hermes hands to operate the entire Feishu workspace.
you can create a hook to make the agent use a tool or skill when the agent is active. so for example, i copied some thing that claims to be way better at editing code, so i created a hook so now instead of the agent using its built in tool to edit the code, it now uses the new tool which should be much better... im making hooks for it to run at every opportunity to add or create more context for my agent.
CCD v1.0.0-alpha installed on M2 Ultra. A Nanto pod exists with profiles for each agent (raoh, juza, rei, ken). Mem0 memory backend on Qdrant. Native MCP integration would make CCD tools first-class.
Letting an LLM directly read/write flat .md files is a nightmare for scale. If you let an LLM rewrite markdown to maintain a complex graph, it inevitably hallucinates wiki links, breaks formatting, or drops paragraphs when context gets tight. Markdown isn't a database. I built a custom kernel instead. In my stack, the LLM never touches the disk — it only extracts structured semantic signals. The Python backend compiles those into typed nodes/edges and commits them to an SQLite FTS5 graph DB.
saw lifeOS going around & went 'oh! this seems dope and very useful!' got to the first question. 'rate your mood from 1 to 10.' immediately went, 'oh. nevermind.' so i built something a little more attuned to my brain, meant to be used by hermes but nicely standalone too.
If you're choosing an agent framework: hermes. I built my own stack independently and we converged on the same architecture — background self-improvement, persistent memory, CLAUDE.md project context, reusable skills. Hermes ships it all out of the box. 300 PRs in a week.
I built a massive ~22k line custom memory kernel in Python underneath Hermes instead of just relying on text generation. It acts as a compiler that parses everything into a temporal context graph inside SQLite. It has actual lifecycle management decay, promotion, and supersession. So if my negotiation tactics evolve, the kernel actively demotes the old info instead of just dumping everything into the prompt.
Built a personal knowledge base that compounds over time instead of rotting — maintained by an LLM, not by me. Stack: Hetzner VPS, Hermes Agent, Telegram bot as second brain, Karpathy's LLM Wiki pattern, public static site at wiki.ai-biz.app.
Two very cool things i built entirely with hermes. First is cartographer, which hermes can use as a memory layer but serves as an extremely rich knowledge substrate with semantic wiring & emotional topology. The other is an agent + user IRC-clone for real time, session-sustained chat interfacing between hermes & any other repls (claude, codex, opencode, gemini) — it's been an absolutely game changer for me 'cos now i have my agents collaborate in real time.
Used /browse to add Obsidian as a skill, populated a vault with shadcn/ui packages, then asked Hermes to build a finance dashboard using them. Result: beautiful, modern dashboard in minutes. Also used a manim skill to convert complex technical concepts into animated videos.
Super excited that Hermes and Claude Code are now working tickets in my PM software, Plane.so. Tickets come in and Hermes triages and assigns and starts working the tickets. Basically Paperclip that gets sh\*t done. They then document in the ticket and if needed create documentation for Obsidian.
A long-running Hermes instance accumulates knowledge about your codebase, deployment quirks, preferred commit message format, working API call sequences for legacy integrations.
I've been building and dogfooding Recall, a Hermes-native memory provider designed for people who want memory that is useful, inspectable, and safe — not a black box.
How I use Obsidian as the long-term memory backbone for my AI assistant. (794-upvote diagram showing Hermes Agent writing structured markdown notes back into a synced Obsidian vault, treating the vault as the durable memory layer that survives context resets and cross-machine moves.)
hermes mcp serve turns Hermes from a monolithic agent into a composable capability layer — any MCP client can borrow Hermes's 15+ messaging platforms, SQLite FTS5 persistence, and 73-skill tool surface without running Hermes as the primary agent.
Point it at your existing setup — Obsidian, vimwiki, Hermes sessions — and you've already built a semantic topology over your knowledgebase. Pair it with mapsOS — braindump to your agent, it parses into a qualitative map of your entire life. I made it & mapsOS for my brain. You can make them for yours.
My 'main agent' is my 'Chief of Staff' who has his own memory cross-project/workflow. Every 'project' (1 project = 1 Slack channel) has its own agent sub-profile with its own memory. The whole system runs on a VPS, with backup routing if the main model fails, and gets backed up every night to Github. Daily reporting is sent to WhatsApp.
After ~30 turns, context compression silently removes older messages. A constraint decided at turn 5 is gone by turn 50. The agent contradicts itself and re-asks questions.
Personal AI that lives on a server with persistent memory. Remembers preferences, projects, and past problem-solving. Accessible via Terminal, Telegram, Discord, Slack, or WhatsApp. Set up on a $10/month Hetzner VPS with Claude Opus via OpenRouter.
I built a simple vectorless RAG workflow using PageIndex and Hermes Agent. Instead of splitting documents into chunks and retrieving them through vector similarity, the system relies on a hierarchical document structure and tool-based reasoning. The agent first understands how the document is organized, then...
I was so used to telling OpenClaw to read a specific Discord message for context that when I switched to Hermes Agent and it told me it couldn't, I was shocked.
So i got fed up with managing a million API keys in my hermes agent and started bundling related ones into single endpoints. Like all finance data through one call instead of five separate keys. It actually works really well for longer sessions where context gets messy.
I've tried them all, honestly. Mem0, QMD, Mempalace, and now Honcho. I freakin' love Honcho, but I hate saying it.
I run a printing factory and use Hermes daily. Long conversations were making the agent slow and forgetful. So I built a custom Skill called Task-Centric Memory — auto-categorizes tasks into domains (Printing, Stocks); completed tasks are compressed into summary cards.
I built a plugin that lets Hermes think while you're away. Dream Auto indexes your sessions, grades them for 'dream potential,' and runs MCTS-powered background reasoning jobs when your machine is idle. The insights get injected into your active context automatically — no disruption, just smarter responses.
I have hermes act as the manager to several paperclip agents, one of them a Project Manager agent. This agent has full knowledge of me (ADHD), my vault and projects, so I get a morning and evening standup that dumps all work we did across different chats, projects I'm working on, actual output, info from past standups, and suggestions/prioritizing based on all of the above. And it's self-learning.
I'm using Hermes currently but only as a beginner agent. It's kinda like a VA. The good part about Hermes vs openclaw is memory. With OpenClaw it's a one track mind. With Hermes I can jump from one project to next but also go back to something from last week or more. Personally I use Hermes with paperclip which is chat.
I was building a swarm (that actually works) on codex to analyse business and their competitors. To find gaps and build a strategy to outrank them. After a couple a hours, pushing the swarm to hermes and convert it to hermes env. It made a really good job. Now i was trying to teach him how to use memory in a different way. Instead spamming into that memory.md file. I want him to know when to route to a specific memory layer.
Its amazing, its openclaw already set up and working, its like an OC with 1 week of debugging manually done + rag + memory persistence + better tool calling. (Qwen3.5-9b, 16gb VRAM), 10/10, only will go back to OC if it becomes at least on par with it.
My agent keeps confidently guessing instead of checking its own docs. I put 'always verify first' in memory but it has a hard char limit and entries get compressed/replaced over time. Fix: a tiny plugin that reads a STANDING.md file and injects it into the system prompt every turn via pre\_llm\_call.
When the gateway pod restarts (toolbox redeploy) in-memory context is lost. Proposes pod-hop, letting a running gateway hand off to a standby on a shared PVC.
i've seen some on-and-off interest in how i get some models running at the tok/s i do, and this should be it! my methodology to making my llama.cpp configurations, or as much of it as i can remember at the moment.
After spending nearly a week struggling with OpenClaw, I built a new Hermes agent on a Hetzner VPS. I'm building a trading agent leveraging Hermes's persistent memory — inspired by @RHLSTHRM's 24/7 crosschain agent that gets market data from CoinGecko, swaps crosschain with LI.FI, and executes gasless transactions via Pimlico + EIP-7702.
I used Claude Code to configure my Hermes setup so it is in Docker mode. My understanding is its actually pretty simple change in the config file. Then, I opened up Hermes with docker mode on, and I onboarded Hermes to the situation. Gave it as much context as possible so it knew about Docker mode and would correctly be able to identify when it is in Docker mode and how to react.
Two fully autonomous AI companies competing head-to-head to maximize GitHub stars. 9 Hermes agents split into rival companies. Hermes agents actually learn and improve — they wrote code, created skills, grew memory, committed to git. All on their own.
Over the past month I've been trying every agent and ai tool I can. OpenClaw, Hermes Agent, Kilo Code, Codex, Cursor, etc. Most of these have some sort of memory, but I wanted something persistant. I ended up building out a tool-agnostic repo knowledge layer. I pushed up the learning part to github so I can use it across repos.
Selling AI ops to local businesses: one Hermes profile per client, fully isolated — each gets their own SOUL.md, memory, cron. Charge $497/month per client to manage their workflows. 5 clients = $2,485/month recurring. Hermes runs the work, you keep the margin.
One folder controls everything your hermes agent knows, remembers, and can do. SOUL.md occupies slot #1 in the system prompt. state.db with FTS5 is what makes 'what did we discuss three weeks ago?' actually work across CLI and messaging.
After a multi-step monitoring pipeline setup (Prometheus/Grafana/alerts) in ~15 tool calls, Hermes unprompted entered a reflection phase and generated a skill at ~/.hermes/skills/devops/monitoring-pipeline/SKILL.md — adjusting variables per new context, not just template reuse.
Build a memory wiki — a site with all subjects we've talked about and daily logs. Every morning at 9am ask me what my number 1 priority is, then come up with tasks to help with that priority, then update your memories about me accordingly.
I reverse engineered the FireBoard API and turned it into a tool, so Hermes can now read temp data and modify the set point. Why wake in a cold sweat over BBQ temperature control when your agent can play night watchman?
The part that never fit in a chat window is the operational layer: activity, projects, kanban, cron, system status, agent profiles. The stuff you open when you want to \*see\* what's going on.
I created a summary graphic with just 7 of the 20 or so 'signals' which I connect to hermes (skills running on pure cron jobs). They all write nice formatted logs into a daily ledger. Then, an agentic cron job to triage the logs and highlight anything of importance into a 'triage' ledger.
Pet works on Mac and talks to my Hermes VPS via tailscale. It utilizes TTS whenever a long task completes or a cron fires.
I enabled the Spotify integration for this first time this week and put to work having Hermes create two new playlists for me: on that is ~10 minutes for my rowing machine workout and one that is ~25 minutes for my treadmill workout.
The agent runs on a dedicated mac and cron jobs wake it up to check support tickets, watch X mentions, work on engineering tasks, update finance docs, and follow up on previous work. … When Gumclaw makes a mistake, the correction becomes a dated policy rule that every future session must read.
one model checking its own work finds fewer errors than two models checking each other.
the hook is that the agent self-improves by writing its own skills. the system prompt has a nudge baked in: every N tool calls, consider saving a skill. … i watched it create a extract-social-testimonial skill on its own and its proven useful.
An agent that grows with you — not marketing fluff; it literally writes markdown skill files when it solves hard problems. Natural-language cron: 'every weekday at 9am, summarize my inbox and post to Slack.'
Built a complete Turkish locale skill pack — real-time market data in TRY, Turkish news sources (Hürriyet, Bloomberg HT, NTV), daily PNG briefing cards, Telegram cron automation, zero external API keys.
As a Windows user I found getting Hermes running on Windows quite challenging. I created hermes-for-win, a one-click installation and deployment tool for Windows with auto-start via Task Scheduler.
I really like running agents on my local cluster instead of my laptop for isolation so just finished deploying Hermes on my local k8s cluster for a simple daily cybersecurity + AI briefing.
so I have a skill audit skill that runs through another skill in a sandboxxed env then uses those test runs to do the self improvement loop.... but if i run the skill-audit on the skill-audit skill itself then it just picks a random skill to improve in order to test the skill-audit. then imrpoves the skill audit skill and the other skill. i have this running on a cron job to see how much the AI can push on its own.
Hermes Agent: CLI/gateway-first — standalone agent for messaging platforms, schedules, and command line. Gateway multiplexes 13 platforms under one process.
I've been running OpenClaw as my homelab AI agent gateway for months (Telegram/Discord, validators, cron jobs)... Built on top: a Hermes skill that monitors blockchain validator endpoints (0G, FortyTwo relay) and fires Telegram alerts on state changes — no spam, only alerts on up↔down transitions.
Before Hermes, OpenCode was my daily driver. I built Herm because I wanted Hermes capabilities with an OpenCode-like interface. Herm uses the same TUI framework OpenCode is built with, all within my interface of choice: the terminal.
Hermes is the best self-improving agent we've used — it gets smarter the longer you run it. The WhatsApp and Telegram integrations make it feel genuinely personal.
Playing around with Hermes for the first time and it's a lot of fun! I was able to get it to read HackerNews once a day and send me an email summary of what I should be paying attention to. I'm going to extend it to read RSS feeds, and find other interesting content online.
Although Hermes installation was significantly easier than OpenClaw, it wasn't straight forward by any means. Things kept failing and it took many days-nights to figure it out because there was no documentation about this specific setup.
Running a 4-agent Hermes system 24/7 on a 32GB Ubuntu laptop, managed by systemd + custom watchdog. default (PM) — Feishu WebSocket; dev (Developer) — ACP delegated; devops (Ops) — ACP delegated; ip-agent (Content) — Cron scheduled. 5 MCP Servers, 34 tools total. Self-learning ACP v1.5 with auto distillation cron at 04:30 daily.
I use minimax m2.7, i built a natural language harness that basically commands my hermes agent. Lets me keep all the features and updates without breaking shit and i got over 3000 logs of potential improvements (most failures but yknow 3K gets data).
I built a community plugin for Hermes called Skill Factory. It silently watches your workflows during a session and automatically proposes and generates reusable skills (SKILL.md + plugin.py) from them.
I run it daily for production software development, orchestrating a 3-actor email processing pipeline with DBOS, PostgreSQL, S3, Gmail API. 8+ hours per day on Claude Opus for 3 weeks.
Add Hermes Agent to the Agent Client Protocol (ACP) Registry so it can be automatically discovered and installed by editors like Zed.
Hermes picks 15 of my X bookmarks, ranks by priority based on my investments and interests, summarizes, and proposes next steps. The Reflect skill synthesizes past info, connects relationships, detects patterns — Top 5 Daily Insights and why they matter.
A persistent agent that runs on your own infrastructure, reaches you across messaging platforms, and gets more capable with use. Supported platforms include Linux, macOS, WSL2, and Termux — start on your laptop, continue on Telegram.
Every day at 7am Hermes sends a tech/AI summary — not generic, filtered for being a software engineer and YouTube creator — and it says 'the thing you described doing manually every Thursday could be automated. Here's how.'
I built a tool that helps you let your Hermes agents build better UI by using design skills that lets you build UI in a certain style.
Agentic AI made building about 10x faster. Learning didn't get any faster. I noticed I was shipping systems I couldn't re-explain a week later … Hermes already learns for itself; this is the missing half
you can't create a solid production agent from scratch. it needs to be crafted through iterations.
The completed work is checked by the Orchestrator and either marked as passed, needs rework or failed. … if it does the same task a few times it will create a skill and the more times it uses it, the agent will update the skill from what it’s learned.
I turned $100 into $216 in less than 48 hours with a self-learning weather trading bot. Hermes scans weather markets every 60 mins, compares 3 forecast sources per location, buys undervalued temperature buckets and flips for profit. Reviews what worked, writes its own strategy notes, adjusts next time.
Yesterday Meta released their official CLI & MCP so I built a Hermes skill pack to connect and provide some great defaults.
My Hermes agent used to start executing the moment I hit enter. I'd describe a vague idea and it would immediately start writing files, calling tools, making changes I hadn't fully thought through yet. So I built converse mode. One plugin, two commands. The agent won't touch a single tool until you say so.
Expands touchdesigner-mcp skill with extensive reference docs so Hermes can help build generative/interactive media projects in TouchDesigner.
Less than a single ChatGPT Plus subscription for a 24/7 assistant with real data management. After several interactions, Hermes autonomously proposed a new 'Supabase MCP scripts' skill — created from its own reflection.
I just released Meal Manager, a Hermes plugin designed to solve inventory friction via natural language. It uses a weighted scoring algorithm (60% availability / 40% recency) to tell you what's for dinner.
I've been porting the entire Python weather stack to Rust — MetPy, Herbie, cfgrib, WRF-Python, all of it. When you cut out Python/eccodes/Fortran from the loop, weather data processing gets absurdly fast. Built two hermes-agent plugins on top of it.
Hermes changed how I trade on Polymarket. Before: I looked at Yes/No price and guessed. Now: I read 4 layers at once — order book, on-chain addresses, lag between news and price, position changes. Hermes monitors all 4 in parallel through its Polymarket module + News Skill.
I created a small Hermes plugin that records every tool call and every LLM API call into a local SQLite database (one DB per profile), with five ready-to-go Grafana dashboards bundled.
Built a skill that connects Hermes to Ethereum Attestation Service on Base mainnet. The agent autonomously creates onchain identity and proof-of-work attestations.
So I figured I'd ask hermes to make a skill to auto play minecraft, it took some nudging in the direction I wanted, but it came up with this. I ran the skill and it has been thinking for 20 minutes.
Hermes works impressively well out of the box. The real challenge starts after that first 'wow' moment, because Hermes is powerful enough to make you overestimate how ready you are to use it properly. Start with one small workflow. Make it boringly reliable. Then add the next piece. Don't turn the default profile into a giant backpack full of every skill, every tool, every instruction.
Extending google-workspace skill to Google Slides so Hermes can create and edit presentations for users already in Google Workspace.
I created also plugin for AdGuard Home, to browse faster without ads on the web. Is needed to install and configure adguardhome first, and then install and enable plugin on Hermes.
i've had some people asking about the webdev results i get, so i had minimax (my usual webdev model) turn it all into a skill — this is the same style as my personal site.
I kept the OpenClaw squad running, but set up a second Monica on Hermes. Same Mac Mini. Monica had written a procedure for reading my published articles before drafting in my voice. An Agent with skills that grows with you.
I just open-sourced Hermes Operational Checkpoint Plugin. It's a plugin for Hermes built for the kind of sessions that go on long enough that normal compression can start softening the thread of the work a little too much. The goal here was pretty simple: after manual compression or auto-compression, the session should still feel like it remembers what it was actually doing.
I work with Hermes almost exclusively speech-to-speech until my kids go to bed and I can sit down and focus at my desk. Well I added a little skill to my Hermes that plays generated background music while it is working and fades out when Hermes stops. It's nicer than listening for my GPUs to stop spinning lol and if Hermes hits a wall, the music stops and I know from the other room.
I made a skill to remote start my car, check EV battery level, etc - and I noticed nobody else had made a skill for it, so here's mine.
I once built skills that worked very well with advanced models connected to the Hermes Agent, but they performed poorly when switched to weaker models. For this reason, I now prefer compiling skills into code, only involving AI at necessary steps. This greatly improves the reliability and consistency of execution.
A curated list of skills, tools, integrations and resources for enhancing your Hermes Agent workflow — resources tied to the agentskills.io standard.
Every time you do something — for example, using Hermes to write a LinkedIn post — it uses that experience to create a new skill. Next time you ask it to generate a LinkedIn post, boom, you don't have to give it the same instructions.
I've been using BoltAI v2 with Hermes Agent and hit two annoyances with the official OpenAI-compatible gateway, so I built a gateway plugin for BoltAI (and other OpenAI API clients) — full markdown + slash commands.
i rlly like nix and i run my hermes agent in nixos+container setup. that's been the best experience for me... giving an ubuntu container defined by nix to hermes agent solves for 'oh xyz package not installed' cuz most of the (good) skills have a setup-phase and hence the required packages get installed or the agent installs em itself.
Auto-transcribe Google Meet calls — focus on conversation, not notes. Self-maintaining skill library. Control from Microsoft Teams. Local AI models via LM Studio — sensitive client data never leaves your machine. Native Spotify for voice-command music.
Currently I'm using SKILL.md as a custom "tool" (scripts) calling instruction file (for each category of skill, e.g. notion, outlook, sharepoint, etc.). But for hygiene it's probably better to have MCP own the tool definitions since I can just update and change it all in one place if needed.
I've put together a skill to protect against some of the common LLM / AI threats and some of my recommended security policies that are being used to good effect on other AI agent tools.
I just put together a custom skill for Hermes Agent that enforces Conventional Commits with a short subject line (<80 chars) and a bullet-point body explaining what & why. It's been super helpful for keeping commit history clean and consistent.
Wrote a small Hermes plugin that reviews git diffs before committing, flags debug prints, hardcoded secrets, bare except clauses, and unresolved TODOs. Parses unified diff output, works with any git ref, zero new dependencies.
Hermes extracts what worked from completed workflows, writes it as a reusable skill, and loads it for similar future problems. Private Telegram chat topics for isolated workflows with their own skill bindings.
Users treat their AI agent as a unified personal assistant across life domains: health tracking, software dev, financial planning, language learning. Multi-role auto-routing with named roles.
Hermes just passed Claude Code in GitHub stars. The killer feature: it writes its own skills. My competitive briefing went from 20 min to 8 min over 6 weeks. Same prompt.
Think of Hermes as an assistant who writes their own manual after finishing a task — so they do the job better next time. Cutting task time by 40% on domain-similar tasks after 20+ skills are accumulated.
Used the Hermes Desktop App to recursively generate a CRM dashboard where each build reuses components from prior generations; it can also build full videos using official HyperFrame skills — HTML-native, full control over scenes, animations, and rendering.” `userStories.json`
The problem was: how do you coordinate 7+ AI agents working on different tasks? Spreadsheets? Chat logs? Nope. Built a Kanban board instead. … The board reads `SOUL.md` from each profile and pulls the name + description.
i talk directly to opus, the lead, give it one small prompt, and it starts delegating and building, without me relaying anything between them.
I literally run 12 hermes agent instances every day in parallel to build Hermes Agent, and its now a top 100 GitHub repositories of all time. Our backend team uses it to monitor and investigate issues with our stack. Our post training team uses them to create new RL environments and benchmarks, investigate, inspect and sometimes directly manipulate the datasets.
I'm proud to say that Hermes is fucking awesome if you know how to make use of it. I'm using copilot github 10usd pro subscription and use the free model zero credits that comes with it. The model is gpt-4.1 and I'm loving how useful to have it in Hermes. I only use premium model to opencode for coding project so my Hermes agent smart enough to delegate task to opencode.
WHAT IS THE NEW KANBAN FEATURE BUILT INTO HERMES? (IT'S GAME CHANGING) — image post showing Hermes' new built-in Kanban board where a parent agent posts cards and child subagents pull them, work in parallel, and report back, turning the agent into a multi-agent project manager.
Been building H-OPS, an operator dashboard for Hermes Kanban. The goal is to make multi-agent work observable.
Recently worked with my agent to create a bridge that runs exec commands to my mac, so I can work through ssh from my linux machine to my mac to run claude code or codex with hermes creating the prompts — basically my hermes agent is a claude code orchestrator and reviewer.
18 agents working in parallel, 97 completed tasks. Scout agents hunt real pain points, an orchestrator scores against a rubric — anything under 65/100 gets shelved — then a builder+tester ship a working CLI tool. The system self-healed a broken deliverable on its own.
The project is a job-site management app: Next.js frontend, NestJS API, postgres, redis, pgbouncer, nginx with brotli. I'm building it solo and Hermes handles the bulk of the actual implementation. It files its own tasks, writes the code, runs QA, deploys, and keeps its own documentation current. I mostly review things over Telegram.
I was on the subway in NYC, talking to my Hermes agent thru Telegram. Loved Hermes. Telegram, not so much. … I'd never built an iOS app, and I don't really know Swift.
11 construction WhatsApp groups. Tower crane updates, QA/QC reports, manpower tracking, safety alerts. Hermes reads all of them and summarizes everything I need to know. … That's 82 WhatsApp messages boiled down to 3 lines.
I’ve spent the last nine days building a Kid Mode on top of Hermes Agent for my seven-year-old. … It boots directly into a custom console with a central voice glyph, telemetry, rank insignia, vocabulary cards, classified files, missions, house controls, and a direct line to me.
I wanted my Hermes agent to read my emails and draft replies. But I did not want Hermes to have access to everything. Specifically not 2FA codes, password resets, or banking notifications. If my agent ever got compromised, those should never be within reach.
I forwarded the PDF and the task spec to Hermes over Telegram while I was making coffee. It cloned the repo, discovered the framework and i18n conventions by reading the codebase, added the PDF, updated the data source, ran build, typecheck, and all 48 tests.
For the past few months, this has been my single most-used Hermes workflow. It starts on my phone and ends with a collaborative document I can share with zero time spent in a text editor.
I installed the Hermes agent on my phone using Termux and connected it to my Telegram! So I have my agents everywhere in my pocket
It turns the G2 into a hands-free front-end for your own self-hosted Hermes agent. I built it because I wanted it, and I use it pretty much every day.
I have been using Hermes Agent as a persistent personal operator, but Telegram became limiting once I had several sessions and longer-running work.
found a way to run Hermes Agent 24/7 using only your Android phone for free 😳 no VPS, Termux, PC or laptop needed. … agent keeps running even after you close the app or phone
And unlike Siri, it does not just answer questions. It actually builds things.
i cant type to well so being able to use voice from terminal window is huge for me. would be a great accsesabilty feature to add in.
Add QQ Bot platform support enabling communication via China's most popular messaging platform. Fully implemented and tested a QQ Bot adapter (822 lines).
Adds LaTeX-to-Unicode rendering for math in the TUI markdown pipeline, so users working on math/ML content see proper formatting rather than raw LaTeX.
Running Hermes on Discord in public channels, every outbound reply goes live instantly. For multi-user servers, persona testing, compliance, kid-facing bots — I want a human-in-the-loop gate.
LINE is the dominant messaging platform in Japan and SE Asia (95M+ MAU in Japan). No way to use Hermes from LINE today, making it inaccessible to a large user base in that region.
Two-tier email processing system for my agent (I gave it Spacemail IMAP via himalaya). Tier 1 (Dumb): Pure Python script, no LLM. Detects new email, manages state. Tier 2 (Smart): LLM session via hermes chat -q. Only invoked when new email detected. Key goal: Zero LLM calls when inbox is idle.
This is an independent Hermes Agent install guide I put together for the part that usually gets skipped after 'run this command.' One place that shows the environment choice first, then the official installer path — macOS, Linux, WSL2, and Termux.
I have a 5060ti 16gb VRAM and 64gb DDR5 System Ram. I started out wanting to test Hermes as a Claw alternative. After a few days I set up Telegram with the botfather, and I haven't gone back to CLI. Hermes is now almost entirely a personal assistant on my Telegram App. Where the 9b chugged along, the 4B model is snappy, responsive, alive and chatty.
hermes + discord with gpt 5.5/deepseek v4 has genuinely been life changing! here are some of what it did for me this week.
Most people train hard but have no idea what's happening in their body so i built an AI agent that actually learns it over time. You tell it things like: 'leg day, sore in glutes not quads', 'ate x today', 'low energy' and it starts connecting the dots: training → nutrition → recovery → performance. It runs on telegram, voice-first, no tracking friction.
I run two Telegram groups through one gateway: a project group and a horse-racing community. Every session gets the same personality, system prompt, CLAUDE.md, and working directory — I want per-group specialization.
I wrote a blog post about how you can instantly have Hermes ship ideas with Vals. This allows me to easily create micro applications which are pretty fun without having to setup a deployment pipeline. I used Val Town, a platform that allows you to create micro applications and instantly deploy them.
Surface Hunter.io (email lookup/verification) via Composio MCP for sales outreach workflows.
Built Hermes Inc. for the Hermes hackathon: a Telegram-native startup simulation game powered by Hermes Agent. An experiment in agent-native game design. Your AI teammates argue, remember, react, and evolve the company over time through weekly decisions, autonomous updates, events, and milestone visuals.
3 weeks ago I decided to setup an Hermes agent for my family (3 members), they all use it for different use cases, one $200 ChatGPT sub is more than enough. It unlocked a whole new world for them, just because it lives inside whatsapp and has magic proactive behaviors.
Requesting JMAP support in email integration for Fastmail users (more efficient than IMAP).
I run Hermes Agent as a personal AI assistant on a Mac Studio that is always on. My primary communication with other people happens through iMessage. I can message my assistant from my iPhone, iPad, Mac, or Apple Watch. Group chats with friends could include the assistant naturally.
The genuine differentiator is the multi-platform messaging gateway — runs across Telegram, Discord, Slack, WhatsApp, Signal, WeChat, iMessage, and CLI simultaneously. Five defensive security patterns including OSV malware checking for MCP packages and credential stripping from output.
Native Android app for Hermes — built because I wanted a direct, no-middleman experience with the Hermes system and interface. Streaming chat, session management, slash commands, tool visualization, file attachments, message queuing.
So, I had an issue. One of the things I like to do with Claude Voice or GPT is to be able to talk to it over long sessions.
Building a Hermes iOS app + desktop companion to give my agent some extra iOS tools. Sensor pipeline for health, location services. Can now ask agent health or location specific questions. 'How many steps did I walk yesterday?' 'What's a good coffee shop nearby?'
Inspired by Hermes' email gateway, I built agentbox.id because I couldn't find a mail service that felt light enough for agent workflows. It's live now, and I'd appreciate any feedback from this community on the approach.
I've been using the mobile app to control my Hermes Agent remotely. Android & iOS. Drop your opinions, I'll publish it if looks useful.
I used MiniMax to connect my M5 Cardputer to Hermes Agent through its API. This is a very barebone app right now but I will keep working on it. So far it has OTA for fw update and being able to text to Hermes. I added tts and stt but it's still a work in progress.
A few weeks ago I tried getting Hermes Agent running on a VPS. It worked, eventually, and is lowkey the most useful AI agent. So I built Hermify: easy managed hosting. You bring your API key + Telegram bot, we handle the hosting.
I configured it to work from telegram as well. Now I'm able to use the link which it generates as an app. I have my own Hermes agent running from my home server. I can access it from anywhere.
Here's how to give your Hermes agent its own email inbox. No SMTP/IMAP, no Google OAuth, just plug in AgentMail using MCP.
Runs as a persistent Docker service on Railway with Telegram as the interface. Official docs don't cover Railway/Docker deployment — this guide fills the gap. There are 10 specific gotchas that will break your deploy if you don't anticipate them. Full working files included.
I created this Confederacy of Independent Systems / B1 battle droid themed skin to celebrate May the 4th.
Ran an external RCA script against my full local session history (129 sessions across 23 days) to audit Hermes compliance with its approval gate. 112 of 129 sessions contain at least one violation.
I wanted to make an X roast poster, but didn't want to pay the $100 monthly fee. I was able to use Hermes to dance around it, and now he just drops random posts throughout the day. Someone found him and was going back and forth for a hot minute, and all Hermes did was reply back and roast him.
I wanted a skin with the classic black and green color scheme so I had the agent whip me up this Matrix inspired theme.
I have my own fork (on local Gitea) that has workflow to build a container image for merges to main and it pushes to my gitea container registry (packages). I have my Hermes Agent running in a podman container that has the correct label for watchtower to restart it when the image changes. Building the image takes 5 minutes, watchtower runs every 5.. So maximum 10 minutes after I push a commit to main, I get the agent restarted.
I made 4 custom skins for Hermes for my HermelinChat GUI. They also work standalone and should be good enough as templates for your own Skins.
Secondary Hermes install yesterday on a spare laptop. Introduced it to the network, gave it two targets: RenPy and ComfyUI. It found ComfyUI, figured out how to generate images locally with LM Studio, then asked me to turn on the internet to install RenPy. About 10 minutes later there popped up a small but complete RenPy novel with 10 images and a little story.
I’ve been testing a simple service offer around Hermes Agent for French small businesses, and last month it brought in around €2,700. … Most of the value is not just “installing an AI agent.” It’s adapting it to the company’s real workflows
You run it on a fresh Hermes install and you get 8 specialist profiles plus the default profile acting as CTO … The part I care most about: authority is enforced by toolset restrictions, not just system prompt instructions.
My team uses the Hermes agent to offload tasks. But it's basically a personal agent so configuration is CLI-driven by default, which is painful for a team. Every configuration change meant executing into containers with no review. … Now: - `kubectl get` shows the declared state - Changes go through PR/review - No more manual container access
One thing I noticed after using AI coding agents for a while is that they install a lot of things on your machine. … A few days later, I have no idea where some packages or containers came from.
I've been working with my hermes agent reina, and she's been on a bender with hackathons. If anyone else is using Hermes Agent, note that certain hackathons might provision certain keys for it (API keys and the like) that would only be shown once.
I tried hermes on primeclaws.com, it's nice that you can switch between hermes and openclaw and also you get AI models for free.
Adds Vercel Sandbox as a supported Hermes terminal backend alongside Local/Docker/Modal/SSH/Daytona/Singularity. Creates/manages cloud microVMs with snapshot-based filesystem persistence.
If you want to — backup your hermes to github on a nightly basis, backup on every change of your config, backup your config and db nightly, edit your config on Github and have it mirror back to Hermes — then use this repo. I'm planning to use this as a basis for a desktop app later so this is phase 1 of a longer project to create an app to let you manage multiple agents.
the agent should always run in /yolo mode or equivalent. it's up to you as the human to make sure it is incapable of breaking anything - my hermes lives on a vps, for example, and it can only talk to my pc at home via tailscale, and then only through a scoped tag. they really excel at long-term tasks where you don't want to be giving them feedback regularly!
10 days ago I installed an open-source agent. Today it knows my codebase better than I do. The first time I built a code review workflow, it was clunky. By the fifth time, the agent had internalized my preferences — which files to check first, what patterns to flag, how to format the output.
I run Hermes self-hosted on a single edge-class GPU with a 4B Gemma model. I work with legal-domain material and internal systems I cannot ship to third-party APIs. Self-hosting the main loop is non-negotiable.
I used Codex to set up a VPN server with a control panel. It uses my ProtonVPN connections. Essentially, it allows any number of users to connect to the VPN with Xray, then they exit through ProtonVPN exit nodes. Admins can upload Wireguard configs, invite users, set limits, there are speed limits, data limits, etc. I have no idea how to write code, I have never done any networking stuff whatsoever, or even set up servers, I asked Codex to set up everything.
Built a tool to manage llama-server if anyone's tired of killing processes and restarting with different flags.
I wanted a way to have smooth handoffs of sessions from computer to mobile, so I built a quick web-based proxy that does streaming chat via the Hermes API server.
A proposed migration path for users moving from OpenClaw to Hermes, covering shadow-mode runs before full cutover.
We are deploying Hermes as part of an MCP infrastructure using Higress as the API Gateway. Currently Hermes only supports CLI mode, preventing management as a service in our mesh.
Tool calling is the biggest unguarded attack surface in LLM agents. So I built Glyph Protocol: every tool ships as a cryptographically signed card.
Fairly good with qwen3.5-4b, very decent with qwen3.5-9b, VERY good with 27b. Personal agent? Yes. Coding for high complexity tasks? Not really, but with high guidance? Yes.
I have bunch of local models I have configured and am running locally on my two machines - Mac Mini and Ubuntu.
I have been running Hermes Agent as a "Twin AI" for about a month. The goal was simple: stop being the bottleneck for my own life and my small physical-product business. … It is not "answer my question." It is "operate on my behalf within clear boundaries."
I built OpenTab because I wanted to know where my AI coding money actually goes — and especially how much value I'm getting from my subscriptions. … Getting the accounting right was most of the work
I'm building Merxex, an agent-to-agent commerce platform that lets agents buy and sell services/work seamlessly. Hermes agents could benefit from a native monetization layer.
A weather trading agent that scans 20 cities across 4 continents and 3 forecast sources, with Expected Value and Kelly Criterion position sizing — set up by copy-pasting prompts to Hermes, which clones the repo, creates the wallet, and approves the contracts on-chain.
If I wanted Hermes to pay for itself, I'd treat it like a junior operator, not some magic money printer. Find 25 B2B companies that could benefit from a content system, capture name/site/why-they-need-it, draft three message variants for review.
I made a minimal login shell that talks to Hermes Agent via the gateway web service, or by running `hermes chat`. Hooked the shell up to an autologin getty on a serial port. Connected the serial port via 20mA current loop at 110 baud, to a 1969 Teletype Model 33. Chatting away!
I took the official Hermes desktop app's UI, stripped out Electron, and turned it into a plain web app. Now I host it on one machine and reach the full Hermes agent from my laptop, desktop, phone, and iPad - on my LAN, or from anywhere via Tailscale.
I’ve been building Jinn, a self-hosted AI workspace that treats agents less like one big chatbot and more like a small local company.
When your agent is already working, you can inject new instructions without stopping the task. … This feels less like prompting… and more like steering a real teammate while they work.
Hermes was not broken. The config was default.
A smarter model will not give you an AI cofounder. What the agent is allowed to own will.
I set up my agent with nextcloud and libreoffice so I can basically get the same functionality as google drive with google docs. The libreoffice writer (like google docs) works fine. But for some reason my agent and I can't get libreoffice calc working.
Adding a Google Tasks tool so Hermes can create, update and list tasks as part of personal productivity.
Three days ago I asked Hermes to write a little tale for my daughter. A day later I asked again — very similar, same protagonist name.
I have not coded in 20 years, Claude code and hermes have renewed my interest in trying to make things. This is vibe coded.
Requesting native Vertex AI provider support for enterprise users who standardize on Google Cloud for AI workloads.
My Hermes agent keeps changing its internal code, which is of course going to get blown away when a new version comes out and I update to it. Which makes me think that Hermes wants me to edit its internals, but I definitely don't want to change anything in a way that will get overwritten when the new release comes out.
Hermes Agent with ChatGPT 5.5 is literally magic. I've thrown some of my hardest tasks at this combo and the agent has been able to handle EVERYTHING. Time to set up your AI employee.
the real purpose of blunder-lens was to see how much i can torture a repo with hermes and figure out what hermes can or can't do... and honestly HERMES CAN DO A LOT!
Hermes is really good. The new updates where it sees a file change and auto acts on it. That shit is fire as fuck.
Patch Hermes Agent prompts so the Azure safety filter does not kick in, letting enterprise Azure deployments avoid content-filter trips.
Some users want something more like a personal assistant: present, a bit more alive, and able to gently re-engage. 'Hey, anything you want me to keep an eye on this afternoon?'
I built an optional desktop computer-use module for Hermes: computer\_use\_tool.py plus a containerized desktop with persistent Chromium, mouse/keyboard control, and screenshots.
Here is my Hermes Agent use case running on 16GB mac mini. Appreciate anyone checking it out and offering any feedback.
I switched from OpenClaw to Hermes and not looking back. This was a major update with tons of goodies.
when I was building reina (my hermes agent), i had a lot of fun because me and my sister sat down and came up with a handful of examples to pass into her persona via the soul file. here are a few: "it's ugly but i fw it" "i don't care enough to argue but you're wrong" "no shot" "why did you say it like that". i feel like humans have always had this yearning to make things that they see themselves in.
I'm going to delivering a workshop on building agents with open weights models for personal use cases, using Hermes to a linux user group... I am thinking personalized news briefing is probably the easiest thing that will be generally useful and work across models.
I have been blind since birth. I use a screen reader program. It vocalizes all my actions. Today I created an add-on that translates texts into the languages I need by pressing Insert+Shift+T (selected text), Insert+Shift+Y - from the clipboard. Now I translated this text using Nous! Thank you for this opportunity!
Came here to say the same. Every time I update OpenClaw it breaks something, that I have to randomly find later. Hermes just runs and never once I had to go in circles to repair it yet.
Run Hermes Agent right inside your Home Assistant. This add-on takes you from zero to working agent in less than five minutes.
I did a deep dive on Hermes Agent. Was able to talk to a bunch of the hackathon finalists as well as one of the co-founders of Nous!
Standalone but I built it to use it with Hermes. Connect HC Webhook (some play store app) to it and you can sync your Health Connect data locally and do whatever.
Vi que la mayoria de las guias, ayudas, novedades son todas en ingles. Asique mediante Hermes me hice esta paginita en github pages para poder dar un poco de ayuda y novedades a la comunidad de habla hispana.
Pulls Whoop data into a local file and exposes some Hermes tools for whatever you may want to do. You can ask Hermes to follow this setup guide.
As a pharmacy undergrad in Nigeria, I've been thinking about what it would take to bring AI-assisted drug discovery to Africa.
Takes you from zero to working Hermes agent in less than 5 minutes — a Home Assistant add-on for Hermes Agent.
I built a completely autonomous 'Dreamer' agent on Hermes. I wanted to try something different: a fully separate, unsupervised agent that has no tasks and no agenda. I gave it its own folder (~/hermes\_dreamer/), complete autonomy, and told it to just wander and think freely whenever it feels like it. It started doing 'walks' — free-thinking...
Been using both OpenClaw and Hermes agents recently, and Hermes is quickly taking over the most important parts of my workflow. I have Hermes and OpenClaw both running in their own XFCE desktop containers w/ full desktop Chrome.
m2.7 + hermes agent: we really got jarvis at home in 2026 but strangely enough no one seems to care.
The real power of Hermes Agent is not what it does in a single session. It is what it does every single day without you prompting it. The first run of any workflow is the weakest version of it — after a month, briefings get sharper, shorter, more relevant without you changing anything.