1. No, my dude, I am not “tired of entering passwords.” Entering passwords has been solved securely and effectively and I no longer think twice about it. What I am tired of is waiting for the magic link to arrive, and I am super sick of the havoc that wreaks trying to just do a simple Oauth grant. 😤

  2. LocalGPT is a little oddly named, to the extent the “local” part is something you can aspire to if the hardware economics of a local LLM make sense to you. But the interesting part is more that it’s very “OpenClaw lite.” Still has the run of the platter, etc. but it’s a single Rust binary and feels a bit less sprawling. It also has a little less personality out of the box, and it feels snappier.

    Comes with a built-in CLI chat interface, a built-in webserver on localhost for web-based chat, and there’s a Telegram interface that hasn’t been attached to a tag yet.

    Curiously, it is also less, erm, “aware” of its own existence as a mediating layer between it and whatever LLM it is connected to. OpenClaw handles questions about itself with operating information that it is OpenClaw you are talking to. LocalGPT periodically answers questions directed to “you” in Claude Code’s voice.

    I like that, by virtue of running it through the Claude Code CLI, it has access to all the tools Claude Code has: Any cloud MCPs and skills pass through to LocalGPT, so it shares basic-memory with any other LLM talking to that MCP along with its own memory layer (which is fast and transparent).

    It is far less likely to induce the kinds of quasi-religious experiences OpenClaw was inducing because it feels more stripped down and the seams show more. That makes it a way better experimental tool to me. There’s less complexity, it’s more transparent, it’s lighter weight, and it’s less aggressive. At the same time, it borrows some of the patterns that make OpenClaw more useful out-of-the-box than the kinds of Jarvises you see people cobbling together.

    As with all these things, it is reflective of the discipline that produced it, so I have to steer it toward my “humanist manager” use cases more. But the transparency, speed, and simplicity mean a little less “what on earth is going on in there” and a little more “if I’m not sure, there’s a limited set of Markdown files this thing is using to ‘know’ stuff, so I can introspect and nudge.”

    I fed it a bunch of blog posts about management and work, along with a couple of skills assessments, and it has become a good planning aid because it nudges me to remember the things that matter to me during a time when it’d be way easier to embrace my inner asshole.

    Keep reading →
  3. On a more prosaic life note: “Get up at 5 a.m. and conquer” is basically a suicide pact for me; but “get up at 6, do little with a screen, get around to breakfast at 7, organize your day by 8” is pretty mellow and sustainable.

  4. I started to write about OpenClaw and got as far as “the story so far on my past attempts to do some of what this is packaging up” before realizing there’s a certain kind of writing about tech that borders on making someone spend brunch listening to a meticulous retelling of last night’s stress dream. Since this blog is as much about me jotting down a few things for future reference as it is the benefit of others, I’m glad I stopped typing so I can just drop in a limited observation:

    “The Gap Through Which We Praise the Machine” remains the best expression of what I keep learning with each wave of “no, it’s different this time” with this stuff: OpenClaw addresses things I was trying to cobble together for myself in a coherent, intelligent, useful way that I personally do not have the wherewithal to build for myself; but the underlying tech on which it is built still requires scaffolding, kludges, and computationally expensive workarounds, and you still end up having to do work to get it to work.

    It is a definite step forward. It’s a proof of concept that is a few steps closer to what I think a lot of us imagine when we talk about the possibilities of agentic AI. I really appreciate that it is a bit more proactive in its way around proposing and implementing tools that reduce the burden on me to craft hyper-specific prompts or continuously redirect it. But you’re still doing work to get it to work.

    The religious rapture with which it is being greeted is not a comment on the technology. It is a comment on the narcissism the tech industry is goading us all into with these things.


    My OpenClaw instance is operating in a personal Slack account, btw. I have a few channels set up where it provides slightly different kinds of assistance depending on context and the kind of help I need. I picked Slack because the integration picture is good with several tools I use. If OpenClaw can’t solve a problem with a tool it made, or if I need to make a quick update to state on a given piece of information, it’s a /command away with the appropriate integration. As I get ready to start a bake-off on agentic AI tools at work, this is helpful validation for a few ideas I have about where Slack fits into it all.


    I gave Chrome’s new “run the browser with Gemini” feature a try with a pretty mundane task: Starting from my Amazon Kindle account, find and remove the samples from my library of 400+ Kindle documents.

    It was an exercise in utter misery. Amazingly bad. I made it stop after watching it “think” for two minutes just to ask me four times before it would remove the first sample, at which point it said “well, this is a bad website so it’s understandable this was not a good idea.”

    I mean, who knows what Amazon’s web team has done in its quest to thwart a few common workflows people use to automate Kindle management (and de-DRMing). It probably had a point.

    “Let’s see in a year!”

    Keep reading →
  5. Our camper got broken into. Filing a report I forgot the plate #, so I typed “camper” into Apple Photos search, found a picture with the plate & copied the text-recognized #. It took less than 10 seconds from “search” to “paste into the report” from 43k photos.