Item — HN Reader

Openclaw has been a gamechanger for me and my friends. I invited it to a groupchat and my friends are enjoying it a lot. It has analyzed the whole group conversation (nothing sensitive), built a personality for each individual user and noted down how everyone talks and what their interest are and their relationship to each other. It has also started to mimic the way we all speak, it barely feels like its an agent in the groupchat anymore. It helps us plan, discuss and roast each other.

A couple of things I have done to my Openclaw instance is the following:

- It runs in docker with limited scope, user group and permissions (I know docker shouldn’t be seen as security measure, I’m thinking of putting this in Vagrant instead but I don’t know if my Pi can handle it)

- It has a kill switch accessible anywhere through Tailscale, one call and the whole docker instance is shut down

- It only triggers on mentions in groupchat, otherwise it would eat up my api usage

- No access to skills, has to be manually added

- It is not exposed to wan and has limited lan access, runs locally and only communicates with whatsapp, z.ai or brave search

With all those measures set, Openclaw has been a fantastic assistant for me and my friends. Whatever all those markdown file does (SOUL, IDENTITY, MEMORIES), it has made the agent act, behave and communicate in a human like manner, it has almost blurred the line for me.

I think this is the key to what made Openclaw so good https://lucumr.pocoo.org/2026/1/31/pi

What’s even more impressive is that the heartbeat it runs time-to-time (every halv an hour?) improves it in the background without me thinking of it, its so cool.

Also, I am so thankful for the subscription at z.ai, that Christmas deal was such a steal, without it, this wouldn’t be possible with the little budget I have. I’ve burned over 20m tokens in 2 days!!!

4 comments

Could you elaborate more on what you find useful about it? I'm struggling to think of a time where an assistant would have been useful in any chat I've been in, but this seems like you've put a lot of effort into it so it must be doing something for you

I wouldn’t say it has contributed to my quality of life or anything, its more of a fascinating agent to play around with. In terms of social acceptance it has nailed it, that is what makes it so fun to use together.

It could probably become useful if I connected my email, calendar, contacts, a browser and exposed devices/services at home to it. But I am too afraid of doing that to be honest.

Remember, it wakes up every half an hour to see if it has any chores. My fear is it would do something stupid that would affect someone else during the night when I’m asleep.

Given how 'simple' it is, I'm very impressed. With it, we built www.clawx.ai over the weekend, helped fix/repair other projects like interface.supernal.ai (used in testing), launched www.computationrights.org. All while I was doing family things.

It certainly is not perfect. Not completely autonomous. It is much faster than Claude or Cursor.

It's already getting better. We're re-architecting its memory, pulling in research, notes from open source, trial-error it ourselves too.

backlog.md has been nice to help with organization but it is not sufficient for complexity or multiagents etc.

Would you mind telling me which model and version you’re using and what authentication mechanism? Is it piggybacking your Max/Pro subscription or did you settle for using pay-as-you-go API costs?

I use GLM-4.7 by Z.ai.

For authentication mechanism, I guess you mean when the agent calls for the model? It’s through api keys.

The subscription I have is the coding plan lite (3x usage of the Claude Pro plan), ~7$ / quarter.

They're mentioning using 20M tokens via z.ai subscription. GLM 4.7 is probably the model then.

As a piece of software it’s pretty awful frankly. Here’s an example : memories are written using the write function but the write function truncates files. There is an edit function but it requires substitution and can’t append. There is no append function.

It also really depends heavily on large models. It’s not practical so far to run on anything that fits into a 4090 because the tool calling semantics are complex, and the instructions pretty vaguely grounded by default. It requires a lot of prompt tuning to get to work marginally. Probably with a clearer semantic in the tools and some fine tuning things would be better on this front. I’ve tried a variety of quantized high tool following models and it’s pretty hit or miss. The protocols around heartbeat and stuff are unnecessarily complex and agentic when they could more reasonably be imperatively driven. It seems to depend on token burning for life more or less.

I frequently see sessions get confused in its internals and general flakiness.

It however has made me consider what a system like this might look like with a better hierarchical state machine, management interface, hierarchical graph based memory, etc. It’s too bad I’ve got a day job and a family at a time like this! It’s a fun time in computing IMO.