# you need to use Hermes RIGHT NOW!! (goodbye OpenClaw!!)

[Watch on tuber](https://tuber.guzus.xyz/video/QQEgIo4Juxg/summary) · [YouTube](https://youtu.be/QQEgIo4Juxg)

**Score: 63/100 · skim**

NetworkChuck says he moved all his agents from OpenClaw to Nous Research's Hermes agent after using it for a month. His five reasons are the vibe, more disciplined memory, the team and its mission, skills the agent creates and maintains itself, and reliability that makes it feel like a product rather than a hobby project. He also walks through installing it and building a Harry Potter-themed IT admin agent called "Ron Weasley," with clips from co-founder Jeffrey Quesnelle.

- **0:00-1:41** — **The pitch and the framing.** Chuck is tired of fixing his OpenClaw agents. He notes that Hermes has become the fastest-growing GitHub project and recently passed OpenClaw in OpenRouter token usage. Hermes is the only agent he trusted enough to give his wife, who calls hers "Honey." He says the headline feature is that the agent "grows with you" and gets better by day 30, previews a Hogwarts-themed IT troubleshooting agent, and mentions an OpenClaw migration path.

- **1:41-2:22** — **Reason #1: the vibe and the mission.** Chuck shows the polished Hermes website and contrasts it with "some random lobster project." Quesnelle explains that the team cares about look and feel because people are "creatures of this world," not just terminally online. Nous Research started as hackers on Discord trying to build open-source AI. The aesthetic is a deliberate part of the brand, not decoration.

- **2:22-5:31** — **Installing on a VPS (sponsored by Hostinger).** Hermes is lightweight Python and runs almost anywhere; native Windows support is said to be coming. Chuck uses a Hostinger KVM 2 VPS running Ubuntu, with a coupon code, and says it can run Hermes and OpenClaw side by side for comparison. He SSHes in from a terminal and runs the one-line install command from the Hermes site. The sponsorship is clearly disclosed. While it installs, he explains the inspiration for his IT agent: Nous runs Hermes agents in its own Discord as a sysadmin "team member" that debugs infrastructure problems people report, and that agent has built up a large set of debugging skills on its own.

- **5:36-9:37** — **Setup wizard: model, messaging, service.**
  - **The model ("the brain"):** Hermes is a harness that works with any LLM. Options include local models via LM Studio (Qwen works well), OpenRouter, or OpenAI Codex, which lets you use an existing ChatGPT subscription at no extra cost. A later insert says Grok subscriptions now work too. Chuck logs in with OAuth and picks GPT-5.5.
  - **Terminal backend:** he keeps it local, noting it can also run remotely.
  - **Messaging:** there are fewer platforms than OpenClaw offers, which reflects Nous's philosophy of doing a few integrations well rather than supporting everything. He picks Telegram: create a bot with BotFather (named Ron Weasley, username must end in "bot"), paste the token, then get his user ID from the userinfo bot and allowlist it so the agent ignores everyone else.
  - **Running it:** he installs the gateway as a systemd system service and runs it as root. He openly says this is not best practice and is only acceptable because this throwaway VPS belongs to the agent.
  - He then confirms Ron answers both in the terminal (which lists the built-in skills) and on Telegram.

- **9:37-12:23** — **Reason #2: memory, and building Ron's identity.** At first Hermes looks like any other agent. Chuck gives it a long persona prompt: Ron Weasley, following his father's fascination with Muggle tech, now working in IT. Hermes visibly calls tools and rewrites its SOUL.md file. When Chuck tells it about himself and his team, it writes to `~/.hermes/memories/USER.md`, which holds facts about the user and is curated by the agent. When he describes the environment, it writes to MEMORY.md, which holds technical and environment context. Chuck admits OpenClaw has these same files, and migrating mostly means copying them over. The difference is how Hermes manages them.

- **12:23-14:16** — **Why Hermes memory stays sharp.** Both tools load USER.md and MEMORY.md into the system prompt at the start of each session. Hermes adds two things:
  - **Hard size caps:** about 1,375 characters for USER.md and 2,200 for MEMORY.md. When a file fills up, the agent has to decide what to drop, which keeps the context distilled and focused. Chuck says OpenClaw has similar mechanisms but tends to bloat over time unless you watch it closely.
  - **Mid-session memory checks:** roughly every 10 turns (Chuck isn't sure of the number and tells viewers to have their agent check), a background agent reviews the conversation and decides whether to update memory. Most harnesses, OpenClaw included, only do this at compaction or when a new session starts.

  He credits these two features for OpenClaw feeling clunky and bloated by day 30 while Hermes doesn't.

- **14:16-17:01** — **Optional add-on: Honcho memory.** `hermes memory setup` shows the built-in memory and past-session search, plus plugins. Chuck used Honcho (cloud-hosted; it can also be self-hosted) for a month with a personal agent called "James Potter." Every message is also sent to Honcho, a separate service that reasons over what you say and builds a "peer card" about you. On each turn it injects whatever context the agent needs right then. His dashboard showed conclusions like his daily habits, that his wife Kathy doesn't like spicy food, and an uncomfortably accurate trait: he "gravitates towards tool building" to avoid high-stakes communication. He had real personal breakthroughs on morning walks in Tokyo. Honcho also works with OpenClaw, but he says it's less of a first-class integration there. He raises a fair worry: you're trusting the model to learn about you and correct itself. Quesnelle's answer is "get out of the way of the models," because they're smart enough to work out what you want.

- **17:01-19:18** — **Reason #3: the people and the origin story.** Quesnelle says Hermes Agent started six or seven months before OpenClaw came out, as an internal tool for prototyping recursive self-improvement in model training. When the team tried OpenClaw, it felt "clunkier" than what they already had internally, so they released their own. Nous describes itself as researchers building "humanistic, censorship-free, democratic AI." It also trains its own Hermes models, which are separate from the agent. Chuck argues that who builds AI tools, and why, matters. The full interview is going on his paid NetworkChuck Academy alongside a Hermes course.

- **19:18-23:50** — **Reason #4: self-created skills and the Curator (the main feature).**
  - **Skills Ron wrote himself:** Chuck asks Ron to set up a headless Twingate VPN client into his studio network and then scan the network. Without being asked, Ron runs a self-improvement review and creates a "Twingate client operations" skill. Quesnelle says the system works like a crude model of human learning: struggle through a problem, save what worked as a reusable chunk, and iterate.
  - **Building vs. downloading:** OpenClaw centers on finding skills in a marketplace, while Hermes builds skills from your own work. Hermes also ships a small library curated by the team. One example is a GitHub PR review skill distilled from the founder Teknium's manual review of thousands of PRs.
  - **Security angle:** Chuck recalls OpenClaw's community skill hub being "malware central" early on, along with several CVEs. He says Hermes hasn't had an agent-related security incident so far and credits its curated, keep-it-simple approach. That's his reading, not something he demonstrates.
  - **The Curator:** a newly added background agent that periodically reviews skills, improves them, and moves them through active, stale, and archived states so self-created skills don't pile up forever.
  - **Less tweaking:** Chuck compares it to his Perplexity Computer video. Because Hermes needs less tinkering, he spends his time using it instead of "sharpening the axe." Quesnelle describes the harness as the model's "hands and feet," giving it haptic feedback from the world.

- **23:50-27:11** — **Live demo: Home Assistant, UniFi, and agent controls.** In a separate Telegram session, running in parallel with the terminal session, Chuck enables the built-in Home Assistant skill and gives Ron an IP address and API key. Ron saves this to memory, switches a studio lamp off, turns it back on in blue, and closes motorized blinds, including one in the kitchen that Chuck hadn't mentioned. On the network side, Ron asks for UniFi access, does an inventory, finds he needs a username and password, and creates a "UniFi network operations" skill. Chuck's point is that he onboards Ron like a new hire ("go look at it, figure it out") rather than pre-building every skill. He also mentions sub-agents and delegation, and the option to interrupt, steer, or queue messages while the agent is working; he queues the API key rather than interrupting.

- **27:11-29:23** — **Reason #5: it doesn't break.** OpenClaw degraded for Chuck over time: updates broke it, or it stopped responding. It felt like a "project" that needed an IT person to keep alive. In a month of Hermes, he says he's had no problems he didn't cause himself. His wife uses Honey for homeschooling, meal planning, and running a house with six kids. He says Hermes has roughly the same features as OpenClaw but deliberately leaves out bloat. He adds his own bold, unsupported opinion: current models (Opus 4.7, GPT-5.5) are good enough that the harness is now what matters most, possibly enough for AGI if model progress stopped.

- **29:23-30:57** — **New features and wrap-up.** He shows the Hermes dashboard: skills and plugins, multiple agent profiles, extra models (for example a separate model for research or delegation), and "achievements." He also shows a Kanban task board where you assign tasks to agent profiles, track progress, add comments, and get asked for input when the agent is blocked. His demo task, a Pokémon-card HTML page, stopped when he hit his usage limit. He mentions a computer-use feature in preview. He says Nous doesn't sponsor him and admits being a fanboy. He encourages running OpenClaw and Hermes side by side or trying the migration.

- **30:57-32:36** — **Outro/filler:** his usual closing prayer for viewers, with no technical content.

**Practical takeaways**
1. **Test before you switch:** run Hermes next to OpenClaw on a cheap VPS or a local machine, and use the built-in migration (essentially copying USER.md and MEMORY.md) instead of rebuilding from scratch.
2. **Start cheap on the model:** use your existing ChatGPT subscription through the Codex login, Grok, or a local Qwen model via LM Studio before paying for API access.
3. **Lock down messaging:** always allowlist your Telegram user ID. Running as root, as Chuck did, is only acceptable on a disposable machine dedicated to the agent.
4. **Onboard it like a new hire:** give it credentials and ask it to explore, so it builds skills from real work. Let the Curator prune skills, and use the team-curated library instead of unvetted community skill hubs.
5. **Optional extra memory:** add Honcho for deeper long-term personal context, keeping in mind that it sends every message to a separate service and builds a fairly intimate profile of you.

## Audience signal

- No high-signal audience themes available.