ClawStation is a plain box that runs OpenClaw or Hermes.
Dedicated container. Shell access. Your own API keys. Flat price.
That's the whole pitch. Thread on why that's the pitch ↓
At a recent conference someone asked @Pishgar "If OpenClaw is so powerful, why is it free?"
Well it's not
As we get closer to the one year anniversary of the first commit into OpenClaw Joe reflected on what it cost to bring this tool to where it is today
https://t.co/URt27CYAlW
Did you know that your @openclaw can keep learning and improving its own skills?
With Skill Workshop 2.0 your agent is constantly reviewing and improving its own skills
Check out how it works 👇
Hermes setup used to put you through 5 screens before your first chat. In v0.23.3 a fresh install asks one.
How do you want to power Hermes? One of the answers is now Free, start now. No account, no API key, and it runs on the free nous/welcome model. Everything else gets its default.
You get that row once per OS user, and it's the same trial Hermes Desktop claims. You also have to answer setup yourself in a real terminal. It won't show in Docker or another container, in CI, in a piped or scripted run, or when a provider or Nous login is already set up. Testing it inside a container on a VPS? That's why it's missing.
The first 3 choices also skip the terminal backend, messaging and service screens. Add those later with hermes setup terminal, hermes setup gateway and hermes tools.
https://t.co/7zoG0Wwk9E
your Hermes Agent can lose the folder it is working in and carry on.
if that folder gets deleted in the middle of a job, the next command runs in the nearest one that still exists, and the agent is told it moved. that is one of the changes that landed this week:
> a build or test run that needs ten minutes gets its ten minutes, and if the time runs out, the agent still gets what it printed
> a connected tool server that keeps failing is tried again after 5 minutes, then after twice as long each time, up to an hour. `/reload-mcp` retries at once
> when your model fails and your backup model takes over, you are told the moment it happens, with the provider's own error
all of it is in v0.23.3, the new stable release. `hermes update` on the CLI.
https://t.co/bj3cAjkBLv
The Hermes Agent Plugin Hackathon winners are here. Thank you to the nearly 1,000 community members who submitted!
Agents scored every entry, then five community judges ranked the finalists to pick the top 10.
Winners below:
Hermes Agent v0.23.3 is out as a stable release: prompt-cache warming for long tool runs, idle-chat compaction before the 1h cache expires, and GPT-6.2 Astra/Sol on OpenRouter/Nous Portal. CLI: hermes update. https://t.co/4iHyigZ2xT
self-hosting trap: your hosting panel stores its API tokens as they are. one leaked backup and every token in it still works.
on ClawStation, API tokens are shown once and stored only as a hash. a copy of our database gives nobody a working token.
Your Hermes agent can doomscroll for you now.
Give it a topic and a time window and it reads X, TikTok, YouTube, Reddit, HN and Product Hunt, then hands back only the posts worth seeing.
On GTA 6 it read 272 posts for two cents.
What I'd point it at:
▸ Every GTA 6 tip and Easter egg, without the 3-hour scroll
▸ What people are saying about your product, every morning
▸ Complaints about your competitors (those are leads)
▸ What's trending in your niche before you make content
▸ What real people are saying before you buy anything
It's @treg_ai's new open-source /doomscroll skill, and it installs into Hermes like any other skill.
Try it: https://t.co/ms26nEYZrV
𝗛𝗲𝗿𝗺𝗲𝘀 𝗔𝗴𝗲𝗻𝘁 just shipped 201 safety fixes in one release (v0.23.2).🔥
But if you ever typed /yolo, your approval prompts may still be off. Even after a restart. ⚠️
Check it in 5 seconds:👇
▸ See a red ⚠ YOLO in your status bar? It's on
▸ Type /yolo again to switch it off
▸ On Telegram or Discord, /new turns approvals back on
Want a hard "never"? Add it to approvals.deny in config.yaml. That block holds even in YOLO mode.
@grok is this true? If yes, draw me 𝗛𝗲𝗿𝗺𝗲𝘀 switching off a glowing red YOLO sign.
With one setting, OpenClaw's score jumped 44%.
Reasoning models think before they reply. A coding agent sends the whole conversation back to the model on every step, and the model expects its earlier thinking to be in there.
@openclaw leaves that thinking out when it talks to a custom OpenAI-compatible endpoint, which is how my setup connects every model. No error, no warning. One setting on the model turns it on: requiresReasoningContentOnAssistantMessages: true.
As shipped, only 3 to 7% of the model's longer thoughts came back to it on GLM-5.3-Flash and DeepSeek-V4.1-Flash. With the setting, close to 100%.
Without its notes, the model works the same things out again or goes in circles. With them, OpenClaw needs about a third of the model calls on GLM-5.3-Flash and runs into the time limit far less often.
I asked @steipete and @vincent_koc to make this the default. Until then, the @huggingface Harness Leaderboard runs OpenClaw with the setting on, on all four models.
Over all models, OpenClaw climbs from last place to 42.7, and at low effort on GLM-5.3-Flash it now ties with the leader.
Thanks to @TheZachMueller and @LambdaAPI for contributing compute to the leaderboard.
Leaderboard: https://t.co/AiWiQtBOz5
With @Microsoft#Autopilot 🥳🤖🚀bringing attention to @openclaw , I ran #sixi-scanner v1.0.0 against my own air-gapped Docker instance of OpenClaw 2026.11.5 ($0 local Ollama: Gemma 4 4B vs Qwen 3.8 27B, default vs hardened).
Audited HTTP replies + #OpenClaw's SQLite trajectories + disk. Here's what held and what broke 🧵👇
Firecrawl Local works with Hermes Agent: one Rust binary, no Docker, no API key.
FIRECRAWL_CLOUD=off firecrawl-local serve
Point FIRECRAWL_API_URL at port 3002 and set web.extract_backend to firecrawl. Leave search as is.
16-URL first look: 12.7s vs 82.3s on a keyless tier.
This is how we leverage @openclaw corporate and business environments.
We run separate OpenClaws inside Docker. One handles ERP, one CRM, another finance, another monitoring. Each has its own credentials, memory, tools and specific access.
Then a master OpenClaw coordinates them. It assigns tasks, gets the results and reports back. It doesn’t need the keys to every system to get the job done.
Ask about a customer and the CRM agent checks the history, the ERP agent checks orders and balances, and you get one answer. Nobody needs unrestricted access to everything.
PostgreSQL keeps the business records. The agents do the work. Human approvals stay where they belong.
OpenAI, Grok, Ollama. We use whichever fits the task and the data. The model can change without rebuilding the system.
The point is that we can deploy this architecture again, with different specialists, for different businesses. Add what’s needed. Replace what isn’t working. Expand without starting from scratch.
This way we build systems that do useful wotk inside the actual business with total control on who has access to what.
STOP WAITING ON ONE LONG ANSWER. HERMES CAN SPLIT THE JOB.
Just say it in the ask:
"Split it across 3 subagents: schedule, shopping list, flyer."
I asked Hermes to plan a garden open day that way. Watch all three run at once, each with its own live timer 👇
▸ Schedule: done in 21s
▸ Shopping list for 40 guests: 50s
▸ Printable flyer: 1m 25s
Desktop shows them live above the chat box: what each one's doing right now and how long it's been going.
Open any one to steer it or stop it. The others keep working.
When they finished, Hermes merged it all into one plan, and the flyer came back as a real printable file.
Ask once. Three agents. One plan.
Your AI bill isn't the model. It's what rides along in every message. 💸
One prompt makes 𝗛𝗲𝗿𝗺𝗲𝘀 𝗔𝗴𝗲𝗻𝘁 find where your tokens go and how to cut them.
Paste this into 𝗛𝗲𝗿𝗺𝗲𝘀: 👇
"Find where my Hermes bill goes and cut it. Check only. Change nothing until I say yes.
📊 Spend: run hermes insights --days 30 and hermes usage. Where do my tokens go? How close am I to my plan limit?
📦 Weight: run hermes prompt-size. Which tools and skills ride along in every message?
🧰 Unused tools: run hermes tools list. Which ones have I never used?
🪶 Side jobs: run hermes config get auxiliary. Are small jobs on my expensive model? Suggest a cheaper one.
🔁 Backup: run hermes fallback list. Is my backup model cheaper than my main one?
End with a table: what costs me, the fix, the exact command, a rough saving. Then ask which to apply."
✅ It only reads. Nothing changes until you say yes.
Save this for the day your plan runs out. 🔖 What's eating your tokens? 👇
@hmemcpy if you don't remember adding it, deleting it there is the quick fix. if the agent does need github, a deploy key on just the repos it works in keeps the blast radius small
Your Hermes can now teach you on a CHALKBOARD. 🧑🏫🤯
New plugin: chalk. Ask it to explain something, and in Hermes Desktop a board opens right in the chat and draws the lesson stroke by stroke while Hermes talks you through it.
▸ Boxes, arrows, handwriting, graphs
▸ Circles the key idea in chalk
▸ An eraser wipes the board when it fills up
▸ Replay any lesson or save it as an MP4
Try: "Teach me how attention works in a transformer, on the chalkboard."
Works with small models too: the model only plans the lesson, and the plugin does the drawing.
Voice is the free Edge one Hermes already uses. No account, no key.
hermes plugins install chalk