OpenClaw

Spent two days trying to get from 9.4 to 9.6 hitting my limits a couple of times. It's getting tiresome

No problems, all of my cron jobs carried over.

Just implemented the upgrade. No problems, all of my cron jobs carried over. Great job dev team! Thank you

It’s supported but carrying over my previous whitelist - I had to allow the gpt6 models allowlist in the openclaw. json modelPolicy.allow

Budget checks should reserve spend before dispatch, not inspect it hourly after parallel work has already consumed the allowance.

In our setup (OpenClaw 2026.9.5), text an agent writes before or between tool calls gets delivered as a message in heartbeat and isolated cron runs, not just the final answer. And NO_REPLY only suppresses delivery when the entire reply is exactly that one token. So "restate the request, decide whether to answer, NO_REPLY" goes out in full.

One thing to watch if you go that way: Claude Code brings its own file-based memory (under ~/.claude/projects/...) and will offer to save things there, next to OpenClaw's memory/ folder. Decide early which one is canonical, otherwise facts end up split across two places and neither is complete.

That's also my answer to the "Claude ignores the agent instructions" comment further down: in this setup it follows them.

OpenClaw can run Claude Code as its backend (the claude-cli runtime), and that's the setup I run on: Telegram stays, and the workspace files (AGENTS.md, SOUL.md, USER.md, MEMORY.md) still get injected into the system prompt, so your rules keep applying.

Treat it like a guest and give it focused tasks and then make sure to get rid of Claude as soon as you can.

I’ve noticed that when you select Claude as the model in openclaw that the session runs Claude code and all the subagents run in Claude code. You can tell because Claude doesn’t get affected by the agent instructions and instead follows its Claude code rules.

And it’s a nosy bastard too finding lots of reasons to look through past session data.

That's why openclaw still wins

Open Claw is just insanely good for doing my digital marketing related work. Open Claw + Claude Opus

I got tired of the roller coaster of bugs and just started on a new clean Ubuntu VM, pointed claude at my old repo backup of my OC and said "I want this feature, that feature, and this..." and it started building me from scratch whatever I needed.

No forgetting a day or week later suddenly due to bloat, it is just very efficient.

The part I like is that autonomy is bounded by budget, role separation, checkpoints, and recovery rules instead of just “let the agent keep going.”

I just set up WhatsApp on a new sim and then assigned that number to OpenClaw. Discarded the sim and it still works…

hourly crons eats up too much of my usage

Nop, I can do a lot of things with OpenClaw and Hermes and I control everything

It 100% can.

dots means you're locked into openai. openclaw still lives in your environment, your persistence. openai dots would need its apps, it's connectors, etc. openclaw let's you build all that. digital sovereignty is still a feature of openclaw, but the difference is definitely dwindling. also, you own your agent files. openai can decide at any point to nerf you and the tasks it can do/allows and you have no control.

Its just I used OpenClaw early on when it constantly broke every day but I'm not expecting that from muse

Open claw was too slow for me. Retired.

I upgraded from 7.1-2 to OpenClaw 2, it wasn't straight forward but it worked, the latest release fixed all the issues I had when I upgraded so it should be fine for you to upgrade.

Anything OpenAI-compatible drops in via a custom base URL, so the real choice is metered tokens vs a flat plan, and if the goal is not watching the wallet, flat is the structural fix: one price, no per-token math.

Yes, subscriptions work with OpenClaw.

API costs more per unit but you can see and cap exactly what you spend, and honestly most agent steps don't need a flagship model - routing bulk work to a cheap per-token tier (GLM/Qwen flash) is what keeps the bill sane.

The catch is metering: consumer plans meter with rolling time windows built for interactive use, and an agent that loops tool calls all day drains a fresh reset in under an hour.

What I’ve learned is don’t rely on the GUI and use HTTP or CLI instead. Create a preflight skill to make sure the cluster is stable before running any missions on the local model. I haven’t had time to test it again but using the GUI is ambiguous at best. Loses track of topology and the CLI/HTTP are more authoritative than the GUI.

it is true just install/update and login to the claude cli and then run openclaw configure and choose anthropic cli

yep, Openclaw had solved the compaction/forgetting issue for a while now. I only refresh a session when I'm in need of deep focus on a new subject.

Muse is cool, but the vendor/ecosystem lock-in is serious. Openclaw can take the best ideas Meta implements and replicate them.

There are some things that companies would never try/allow that harnesses like openclaw give you the freedom to experiment with.

Use two replays. First freeze tool results and state, then rerun only the model. Next freeze the model decisions and replay the harness. The first comparison isolates decision variance; the second exposes transport, permission and checkpoint bugs.

It respected two human gates (to approve the research brief and the script) and ran from 11:30 pm till 3:30 am.

QWEN3.6-27B MLX-8bit handled it with zero drops. It respected two human gates (to approve the research brief and the script) and ran from 11:30 pm till 3:30 am.

it’s the most cost effective with a tailscale egress to chromebook connected verizon fios egress

exaclty love my claw i am on You are currently on OpenClaw version 2026.4.2 (build d74a122).

if i have an issue i just claude code or codex cli to fix the the problem on open claw.

People whose claws work fine post rarely, only those who break are vocal. That's not to say that Openclaw is perfect. That's why we encourage all users to submit the issues/PRs to help us identify and fix problems quicker.

I'm curious about what makes ppl want to use this software when all I see are posts that an update broke everything they worked hard on...

I saw they cut Luna 6 costs in half compared to 5.6. it's so great

Telegram/discord delay time is drastically reduced, easier to manage cron, phone app linking, revert message.

And then i will wait for the most stable release again

I had real trouble with upgrades since June. Tried a few times and had to roll back every time.

I have also tried hermes and it works pretty well, but I have a lot of time into OpenClaw customizing it for various functions and nothing else I've tried works as well.

I had to force my way through the update last night. My database is fairly large at 6 GB. For some reason, the update process makes three copies of the database to ram. I had to increase the ram on my VM to 24 GB and then increase the swap for the update to be able to take place. Very odd process, but I finally got there.

I cannot even upgrade.

I have an openclaw-ops skill that will now by default check github issues before we do anything because sure enough most of the time something doesnt work, we hit a bug for which an issue already exists.

I have an openclaw-ops skill that will now by default check github issues before we do anything because sure enough most of the time something doesnt work, we hit a bug for which an issue already exists.

OpenClaw is more open and easy going on.

I've had to build a heartbeat system into them but I love it way more than open claw.

9.6 made me switch to Hermes and I wish I'd done it sooner.

The real fix IMHO would be to have proper plugin hooks to modify selectively each single part of OC’s system prompt. The currently available hooks are not enough as they filter the whole thing.

I had less than 5% of my weekly usage left when I made the switch. It blew me away how much I was able to get done, using all three agents at the same time, and still haven’t exhausted the usage.

So they turned one request into three, and assumed that models used to direct calls will magically understand this new way of doing things instead of floundering and wasting credits. Are they high? Do they think we're all made of money?

So they turned one request into three, and assumed that models used to direct calls will magically understand this new way of doing things instead of floundering and wasting credits.