OpenClaw's new update is MASSIVE
TL;DR
OpenClaw’s biggest new feature is full ClawHub integration in the CLI — Alex demos
clawhub searchlive, shows how OpenClaw can find and install skills in real time, and then immediately turns that into a custom Mission Control “skills tab” with install, analyze, and “make my own version” buttons.Alex’s main workflow insight is: don’t blindly install community skills — vet or rebuild them — he calls skills the biggest current security gap, says ClawHub is scanning more aggressively now, and recommends pasting a skill link into OpenClaw and asking it to analyze or recreate the skill instead of “raw dogging random skills.”
The new
/btwcommand fixes a real OpenClaw pain point: context pollution — Alex says he constantly gets derailed by side questions mid-workflow, and now/btwlets you ask lightweight one-off questions without storing them in context, using tools, or blowing up token costs.Sub-agent controls got much more practical with adjustable thinking and model selection — instead of every worker inheriting the same expensive reasoning settings, you can now lower thinking levels or assign cheaper models like GPT-5.4 mini/nano for scanners, coding helpers, or repetitive tasks to save money and speed things up.
Alex’s broader thesis is that OpenClaw and Claude Code are not competitors in the way people think — he says Claude Code is a great closed-source coding tool, but OpenClaw is a general-purpose open agent that will SSH into machines, build tools, orchestrate workflows, and act from Telegram with far fewer guardrails.
His most practical warning is that OpenClaw suffers from session bloat, especially with cron jobs — Alex says each scheduled task can leave behind session records that balloon context, slow everything down, and increase token usage, and he claims cleaning old sessions sped his setup up dramatically.
The Breakdown
A chaotic start, then Alex declares the update “massive”
The stream opens with Alex’s internet dying, YouTube apparently going live anyway, and him joking that every creator guide says the first three seconds must be “fire” while his audience may have just watched him stare at the screen for five minutes. Once the dust settles, he frames the stream as the biggest OpenClaw update ever and promises live demos, implementation advice, and probably noise complaints from the neighbors.
ClawHub lands inside OpenClaw, and Alex immediately starts hacking on top of it
The first big update is full ClawHub integration: from the CLI you can now search skills in real time with commands like clawhub search image, inspect options, and install them directly. Alex likes the speed but is visibly skeptical of the name and even more skeptical of blindly trusting community skills, calling skills the biggest security gap in OpenClaw right now even though ClawHub is now scanning them more aggressively.
His real skill workflow: inspect first, or have OpenClaw build a safer clone
Instead of one-click installing everything, Alex says his preferred approach is to open the skill listing on the website, feed the link to OpenClaw, and ask, “What do you think of this skill?” or “Build your own version of this.” Then, in classic Alex fashion, he turns that into a live build: a Mission Control tab that pulls popular skills, lets him search, and adds buttons for install, analyze, or recreate-your-own version.
Why he thinks OpenClaw is in a different league from Claude Code
A big chunk of the stream becomes a rant against the idea that Claude Code “killed” OpenClaw. Alex’s argument is simple: Claude Code is a fantastic closed-source coding tool with guardrails, while OpenClaw is a wild, open-ended computer operator that can find a DGX Spark on the network, SSH in, download Qwen, run it, and build an AI lab on top of it without asking for permission a hundred times.
The custom dashboard works, and Alex uses that to make a bigger point
When his coding agent finishes, Alex opens Mission Control and shows a functioning skills UI with search, rankings, installs, and an analyze action that routes code review requests back through Telegram. For him, this is the whole OpenClaw mindset in one moment: a new capability ships in the morning, and by the afternoon you’ve already built your own interface and workflow around it in one prompt.
Hermes, models, and why task completion still matters more than benchmarks
Alex gives a surprisingly balanced update on Hermes: some parts impress him and may even beat OpenClaw, but he says its “compactions” feel like a nuclear bomb, where the agent suddenly forgets who it is and what’s happening. He also compares Claude Opus 4.6 and GPT-5.4 inside OpenClaw, saying GPT-5.4 often feels smarter, faster, and cheaper, but still has a fatal flaw for agentic work: it often just doesn’t finish tasks.
/btw is the sleeper feature because it protects your flow
The second major feature is /btw, which Alex clearly loves because it solves a behavior problem he actually has all day. If he’s deep in a planning conversation and sees a tweet or random idea he wants to ask about, he can now spin up a side question without polluting the current thread’s context, triggering tools, or burning extra tokens.
Sub-agents get cheaper and more controllable, plus one important warning
The next upgrades are adjustable thinking for sub-agents and the ability to assign different models to them, like keeping Claude Opus as the orchestrator while using GPT-5.4 mini or nano for lightweight scans and helper tasks. Alex says that’s a big cost and performance win, then pivots to his strongest practical warning of the stream: cron jobs and saved sessions can silently bloat OpenClaw’s context, slow the whole system down, and drive up costs unless you periodically clean them out.