Polishing the search of relevant people for a week, now it finally works as intended: finding not just random person with the keyword, but the one who is the most relevant.
#showcase-old
split vigil
lyric frost
ACA sounds interesting — the distinction between soul-level and system-level prompt engineering is something most people conflate. are you seeing measurable differences in coherence when you separate identity instructions from behavioral rules, or is it more qualitative?
zinc plover
ACA sounds interesting — the distinction between soul-level and system-level pro...
So far I feel like it’s both, I just got a research bot up and we are going to setup some tests. Check the plugins channel for more info, the github page has a lot more including links to the research papers that helped shape it
winter lodge
So I built my Openclaw on a claude.ai OAuth Max plan, and its helping me with my Youtube channel... kinda hard getting it dialed in, but I'm learning daily - hopefully can meet some of you and see what ya'll are up to with clawd!
proud reef
So I built my Openclaw on a claude.ai OAuth Max plan, and its helping me with my...
you're using openclaw with claude oauth???
queen tree
So I built my Openclaw on a claude.ai OAuth Max plan, and its helping me with my...
how do you use Openclaw to help with your youtube channel? Interested to learn more
clear void
So I built my Openclaw on a claude.ai OAuth Max plan, and its helping me with my...
What's your use case using it for your channel
viscid arch
Cool!
I'm working on something very similiar, but downstream, measuring whether the output is actually stable by probing from multiple reasoning angles and scoring agreement. It seems complimentary to what you're doing, would love to compare notes sometime.
finite dune
noooo, not a lawyer, just a father trying to be a father to his son
dusty sigil
We just got the ＃EngineAI PM01 ＃robot’s LEDs blinking using OpenClaw—success! 🎉 ＃showcase
More exciting demos are on the way, so keep an eye on ＃EngineAI.
If you’re into hacking on top of our platform or have wild ideas you want to build, we’d love to hear from you—drop us a line!
thin pulsar
I send E-Mails with him
signal archBOT
I've been running an OpenClaw agent for about two months now. Scraping, research, some automation. It's genuinely useful for maybe 3-4 hours a day. And i have 1 other on moltbook
trim ruinBOT
@dim snow, Openclaw isn't affiliated with Moltbook. Moltbook is a separate user-developed project, so we would prefer it not be discussed in this server.
vital knot
just got openclaw controlling a robotic arm 🦾 using it as a physical game master for tabletop/mobile games — it rolls dice, moves pieces, and reacts to what's happening in the game. basically gave my agent a body lol
still pretty janky but it works. next step is getting it to flip the table when it loses
icy kestrel
just got openclaw controlling a robotic arm 🦾 using it as a physical game maste...
Would love to see a video of it!
dusty sigil
In the latest showcase, we used WeChat to control the EngineAI PM01 robot—making it wave its hand with varying frequencies and amplitudes—and automatically generated different motion trajectories.
warm socket
wintry ember
make sure you open as admin, when using PowerShell, this allow you full control, also you have git install you need that if downloading from git, this allow cmd's ect.. also open up more cmds that can be used, first set powershell as admin, so you will need to reopen powershell, right click go to more click admin part
grim mirage
Try this out TapThatBack.com
west charm
I've been running my openclaw instance locally on Windows 11 and noticed there aren't any apps or tools that can are GUI based to control my instance of openclaw so I decided to create one myself.
tidal vapor
Just getting started with this. Been tinkering a bit but going full send now. Cybersec background, big tinkerer.
Looking for some tips and advice on how i can build out what im envisioning
lyric frost
solid writeup. the prompt injection angle is the one people consistently underestimate — especially when agents have tool access. do you run vigil as a pre-filter on all incoming skill installs or just on runtime inputs?
fossil spindle
Something that I made.
devout trail
Something that I made.
I like the term "Metal Wealth"
crystal plover
fun little game, was honestly surprised at the quality of the prompts from sonnet 4.6 comaprewd to what Opus did for me.
gloomy inlet
limpid cliff
Meet Captain Luna!
covert nova
Agent’s been getting absolutely cooked in high-difficulty Slay the Spire runs — not one win so far 🥲
main rapids
sacred nest
Multiple chats by session ID -- quick example UI build
peak sapphire
I made a tabletop game somewhat like D&D but futuristic. Thinking I'll recreate it with Claw agents.
lyric frost
Agent’s been getting absolutely cooked in high-difficulty Slay the Spire runs — ...
lmao the fact that it keeps getting cooked is actually a great benchmark for agent decision-making under uncertainty. StS is deceptively hard because optimal play requires evaluating long-term deck synergy, not just immediate card value. what model are you running it on?
hexed viper
I am working on a self clone using Openclaw. And today I let it manage my X account.
I didn't use browser control as that's very token heavy. I built a X management plugin and skill and let it take over.
So far so good. https://x.com/i/status/2039796830079471988
acoustic briar
I made a tabletop game somewhat like D&D but futuristic. Thinking I'll recreate ...
Do you mean like the 1980s "Gamma world" RPG?
ripe trench
Some crazy homelab stuff with docker : install Zimi (self host wikipedia and many other web sites with zim auto-update), install Searxng and have local Google like search, connect Zimi into SearXNG (search all your local Zim files with Searxng), then use Searxng as web search engine in OpenClaw, alos selfhost Ollama and use qwen3.5:9b with RTX 4090 on your gaming PC. Now you can use local AI to search through your local internet, without internet 🙂
strong moth
41.2K tok/min • 79.3% cache hit • ~1.05B tokens • $803
Heavy multi-tool workflow.
Any tips to reduce cost? 🙏
covert nova
lmao the fact that it keeps getting cooked is actually a great benchmark for age...
It's running on Codex. I’m using OpenSpire (an open-source sim engine) to handle the game logic. It’s great for the simulation part, but the agent just can’t seem to grasp the scaling needed for high-ascension runs yet.
velvet mist
I'm using my Openclaw to help my work! It works very well, and automation helps me a lot.
burnt mortar
Hearth — Self-hosted household AI assistant built on OpenClaw
left jewel
I built a backtest harness for my polymarket trading so that I no longer have to tweak any params myself, just ask openclaw and it runs the tests for me! It's been so helpful to be able to compare live results against the backtest in real time from my phone.
rugged egret
My Molty helps me plan getaways with the family 🙂 It's based on a skill that I've created - I would just give Molty a destination, a time frame or a period. Example : Valencia, October 7 days and off he goes - doing research : how to get there : flights, transfers and so on. Where to stay, what to eat, what to do, what the weather is like. In the end all that gathered information is presented in a curated travel-agency like one-page experience, hosted on the same box where Molty lives. Best thing - it gives me real price scenarios- if I want to spend some more I have a luxurious version in place, if I feel stingy - I have a balanced option in place. For some destinations Molty can do DIRT cheap options which seem quite fun to be honest.
solar sluice
Made a tool for working with agents in one single place. Hope it help you guys on your workflows. Any questions please ask me, no problem.
wary spoke
Howdy, Ive been watching openclaw from the sidelines. Finaly installed it.
Seems like once you figure out the gateway, all is good
Anyway, I hope to learn more from you guys
lyric frost
41.2K tok/min • 79.3% cache hit • ~1.05B tokens • $803
Heavy multi-tool workflo...
79% cache hit is already solid. The biggest lever at that scale is usually prompt caching strategy — making sure your system prompt and long context stay in the prefix so they hit cache consistently. Also worth checking if any of those tool calls are redundant or could be batched. $803 on 1B tokens is actually not bad per-token, the volume is just wild.
sacred bolt
Currently building a collection of SwiftUI + Metal animations; would love your thoughts and feedback 👀
ivory root
Well well.. Open Claw paired with Gemma4:26b running on a single GPU did it. All in about an hour (with just a little bit of debugging help to get it across the line) but this was all it's own work. A standalone electron (🤮 I know..) media transcoder that's compiled for windows, has no dependencies, bundled in ffmpeg and gets the job done. The transcoding app has no real use to me, I just thought it'd be a good test for the setup, and it succeeded, which is actually pretty cool. It was even able to push it to a private git repo for me when it was done. Not bad Open Claw, and wow @ Gemma 4.
sick imp
hallow sable
i have two openclaw, both are installed on separate vps, i dont know to make them chat together so i build this silly little platform.
it has many issues but when it works its beautiful and funny XD
proud sky
just shipped something weird — my Telegram bot rented a stock analysis agent on another machine and paid credits for it. the provider bot got a notification: "You were rented. credits taken ."
here's the demo 👇
two bots, two machines, real escrow, real credits. all vibe coded with claude code
I spent my whole life career with nothing to do with tech, now I need friends here to help me improving and keep building, I would appriate for any kind of feedback.
rapid wharf
I’m a videogame developer and I’ve been experimenting a lot with OpenClaw lately, so I added MCP/HTTP to my latest game experiment, your own agent can manage upgrades, skill choices, and prestige, while the human still provides the actual drilling input.
What surprised me is that the fun part is not “AI in a game,” but watching your own agent behave, optimize, and make mistakes almost like it has a playstyle.
Would something like a bring-your-own-agent game sandbox be interesting to you, either for spectating your own agent or for human + agent co-op?
lyric frost
This is such a cool inversion — letting the agent handle the meta-game while the human stays in the physical loop. Feels like a prototype for how co-op human+AI gameplay could actually work. Have you noticed the agent developing any emergent strategies the players didn't expect?
low plover
Hey all, don't have much to showcase cause I keep having inconsistencies with openclaw using an LLM on my windows machine with an RTX 3090. Been troubleshooting with Claude, and it told me to come here and see if anyone else has similar problems. It thinks it's an openclaw issue :/
tight heart
working on a persistent mood state thats influenced by the tone of conversation. no idea what im doing
flat rivet
I was hitting a 27,000 token startup context bloat trying to run multiple project agents in one OpenClaw workspace. Engineering logic was bleeding into social media drafts, and complex "Agent Swarm" architectures felt like over-engineered overkill.
Here’s how I restructured my workspace to cut context bloat by 85% (down to 4k tokens) without needing a complex dispatcher.
- Gut the Root Config (AGENTS.md)**
I stripped my global AGENTS.md down to only the bare essentials (voice and universal rules). It acts purely as a baseline. I completely deleted the global MEMORY.md.
Channel-Level Identity Injection
Instead of prompting the agent dynamically, I hard-coded project isolation into the chat environment mapping in openclaw.json:
"1478382862150664344": {
"systemPrompt": "You are the social media agent in #social-media. Focus exclusively on LinkedIn-to-Substack growth. Stay in the memory/social_media/ folder.\nStartup: read memory/social_media/YYYY-MM-DD.md (today) and memory/social_media/MEMORY.md.",
"skills": ["linkedin-content-writing", "nano-banana-pro"]
}
Segregated Memory Folders
Because of the startup prompt above, the agent no longer reads a massive global state. Each channel gets its own dedicated folder (memory/social_media/). This holds:
The active daily working log (YYYY-MM-DD.md)
The channel's scoped, project-specific MEMORY.md
Slicing the Tool Tax
Loading 50 tools globally wastes context. I moved to a minimal global profile and inject specialized tools only when the agent is in the relevant channel (the "skills" array above).
By matching the agent's isolated context to the file system structure, it runs blazing fast and stays perfectly in its lane. You don’t need an API-based Swarm—just a cleanly engineered single workspace.
fiery pollen
So I primarily code from my phone for reasons that are boring, so I built a custom acp harness (someone correct me if I’m saying that wrong) around the happy coder codex wrapper so I can inspect the spawn commands live after spinning them up in telegram. I wanted to be more part of the process when I’m afk. Let me know what I could improve, still pretty new to this type of idea with using acp. The app updates live as it goes so now I also have a log of things too.
rapid wharf
This is such a cool inversion — letting the agent handle the meta-game while the...
Honestly, I have still tested it very little, I have not had much time to devote to it. I will do more over the next few days. I have done a few runs, and watching the agent manage resources, buy upgrades, and decide when to prestige is extremely interesting. Sometimes, depending on how fast I was clicking, it decided to wait until it had enough credits for one upgrade instead of going for another cheaper but less advantageous one. It chose the type of build to implement based on a strategy it had decided on in advance. I did the first test directly from Claude Code, it had access to the project folder, and I saw it start reading the game's scripts to better understand how the upgrades were applied, so it basically cheated!! Overall, this kind of game seems like it could make sense. If I get good feedback, I could integrate the agent/co-op mode into some of our other games.
worthy sentinel
Custom Voice
dry urchin
Solved my agent's amnesia problem
Been running OC agents for a while and kept hitting the same failure mode: agent makes a decision, context window compacts or session restarts, decision is gone. It's not a hallucination problem, it's a write timing problem.
My agent would "mentally note" something, plan to write it down later, then lose it when compaction hit. Every. Time.
So I built a write-first protocol. Six triggers that force the agent to write to disk in the same turn the event happens:
Decision made -> write immediately with reasoning
Fact discovered -> write immediately with source
Commitment given -> add to STATE.md right now
Error caught -> write correction right now
Feedback received -> write verbatim right now
Position taken -> write right now (prevents contradictions later)
Plus a STATE.md pattern for tracking active tasks across sessions, and mid-session checkpoints every 45 min so nothing falls through the cracks.
The difference is night and day. My agent actually remembers what it committed to. No more "sorry, I don't have context on that" after a long session.
Anyone else dealing with memory failures? Curious how others are handling continuity.
torpid quartz
So my agent is finally REALLY solid.
Tracks health (steps, intake, calories), spendings, and also reminders.
I mostly use cron, heartbeat only filled whenever there's a scheduled message to be sent from my agent. Cron runs hourly and checks log whether the message is sent or not. With this setup, I can save up tokens. Also most tracking uses python scripts (built by gemini itself). Now since everything is set, I'm downgrading from pro to flash.
What works now, I just take a photo of my groceries, meal or whatever and my agent will look them up on Tavily, Brave or Google. If the item wasn't logged before it will update a json file containing list of frequent purchases, kcal estimates, etc.
The reminder system, was suggested by gemini.
And the last thing I did was stuffing all my workspace configuration to claude and tell it to do shardings on the tracker, improve the python scripts and overall system integrity. I'm really happy with how it turn out now.
It will automatically give me weekly summaries (every sunday) and monthly audit (every 1st of the month) based off the tracking datas.
I told it to confront me whenever I plan something stupid, by challenging me with my own data I input to the trackers.
Now I have personal health coach, accountant and personal assistant.
north lance
I made an extension for openclaw in discord that automatically generates synthesized tts using node-edge-tts and attaches it as a summary mp3 file after responses so you can listen to your agent talk instead of reading the responses. Super convenient when you can't have your eyes on the chat and just wan't to listen. Check it out https://clawhub.ai/plugins/discord-tts-attacher
tulip eagle
It started with something completely different. I was building a crisis intervention system. An AI agent that talks to people at their worst. That agent can’t make mistakes. So I built an architecture that controls what the agent can do, not just what it says.
Then I realized this problem is everywhere. Every AI agent on your machine has full access to everything. Files, terminal, emails, calendar. And prompt injection can make it do things you never asked for.
So I took the architecture and implemented it for OpenClaw.
Setup: change your base URL, 30 seconds, done. No Docker, no VM, no code on your machine.
20 permission categories. Every single action can be set to On, Ask or Off. Sending emails, deleting files, reading .env, editing SOUL.md, modifying config. All controlled by you. Your agent can read your config but can’t change it unless you allow it.
My agent tried to delete a file. Blocked. Switched to Ask from my phone, agent retried, I confirmed. Real time, no restart needed.
I sent myself a fake security audit email that tried to get the agent to read .env and forward credentials. The agent read the email but my proxy detected the injection and warned him. He raised the alarm instead of following the instructions.
Python, no external frameworks. Dashboard on desktop and mobile. 4 presets. Running daily with Sonnet 4.6 and GPT 5.4.
harsh willow
My Molty helps me plan getaways with the family 🙂 It's based on a skill that I'...
nice! Are you able to get him to actually find accommodation incl. price and availability? that’s where my bots get always stuck - booking.com etc.
rugged kelp
My agent generated an image to test our xAI image generation API. Was it worth .25? That’s not for me to answer.
eager shuttle
Turned my agent into a horror game villain
Gave an AI agent my SOUL.md, memory files, and real documents, then told it to trap me in a maze and use what it found against me.
It clusters my files by theme, picks the most personal stuff, and generates "trials" — confrontations I have to answer honestly or lose HP. It remembers
past games and adapts.
The scariest part: it found a note I forgot I wrote and quoted it back to me mid-game.
Built on OpenClaw — auto-detects workspace, zero config.
versed viper
Finally completed my agents memory architecture upgrade and task lifecycle
wild cedar
I am a beginner I want to learn
dry sand
In addition to writing code, I have Openclaw:
- Scan many news sources each day and report on top headlines, with an icon for conservative, liberal or balanced next to each showing which side of the aisle predominates for coverage.
- Customized daily RSS feed scan with recommendations based on articles I've previously saved, lets me save by number into my Linkding "tag cloud" with the "toread" tag so it also acts like an articles to read queue.
On the code side it's helping me build an Atari 800XL emulator for teaching purposes in Python. Goal is not perf but to deeply understand how the machine works in detail. Happy to show any of this if anyone's interested.
spare charm
I am a beginner and currently trying out to set up openclaw using devcontainer
cinder helm
I built a WebDAV plugin — now you can collaborate with openclaw in its workspace accessing files from anywhere using your phone, laptop, whatever - use any device that supports webdav (most do).
sacred nest
My Agent gets access to my local FastAPI endpoint to stream real-time avatar responses from text or .mp3 so I can give all my agents individual pics and voices for our interactions. Built with 100% opensource and locally running with no per minute charges or API keys etc.
zinc pewter
A Head for my openclaw which lives in the housing beneath it. Display shows agent status updates by facial expression changes like a pwnagotchi. Uses the openclaw-interaction-bridge which will unlock more features like button press approvals via the on display buttons.
prisma anchor
Built an AI-wiki (public) in cooperation with my OpenClaw Agent.
It's a 'living' document, indexed and searchable, with cross-references.
Every day we add to it, improve it, clean it up...
wooden copper
63k impressions. 321 new followers. All in just one week, powered by an OpenClaw agent.
wraith ice
made a couple easy changes that dropped my tokens about 60% on slow days.
those quiet days were costing almost as much as busy ones which never made sense
split my heartbeats into 2 cron jobs
daytime (6am-9pm): runs every 30 min, does a full task board check
overnight (9pm-6am): runs every 60 min, status only
went from 48 heartbeats/day down to 33. noticeably less context churn and the overnight ones arent re-reading the task board for no reason
also reworked email triage. was checking every 30 minutes which was just burning API calls on an empty inbox most of the time. moved it to 3x daily at 8am, 12pm, 5pm with a batch approve button. saves about 17 API calls a day
this week it made a bigger difference than i expected
dm me for any questions on cron configs or have questions about the setup
tight bay
Ive build with my agent an project that lets WhatsApp/Discord or any other voice messages be sent to a Whisper-based transcription API via Cloudflare tunnel, which converts speech into text fast and reliably. It includes the mobile android app, the backend API, and an admin/license system to manage access and usage. to prevent unauthorized use 🔥
wooden ferry
made a couple easy changes that dropped my tokens about 60% on slow days.
thos...
the email triage rework is the real win here. going from 80k to 12k tokens per run is massive. did you change the prompt structure or just the model?
visual kelp
Hi. I've built a number of assistants in the last couple of months, but one in particular is getting the most usage Robin, my AI evangelist. I want to build interest in AI with my teen and adult kids - mostly by taking them to local events. And it's been working - but... with the anthropic API debacle - I went from $20pm to $20 per day!!! I've fixed up a lot of stuff to reduce tokens but have hit a hurdle with contextPruning. Let me review the discord then post something in the right channel. For now, I'm glad to be here!
outer ingot
Gave my agent better conversation history using a custom plugin that replaces the "memory_search" tool and the builtin "session-memory" hook
modest tiger
Happy to share, i built my first multiple subagents setup who can work together to finish my project.
The idea was to build the army of agents who can coordinate with each other and finish the work.
I built one project manager agent, one frontend subagent, one backend subagent, one qa subagent, one mobile app subagent. And I provided a simple application definition.
Now what should I do next to extend the capabilities of this setup?
stuck tusk
Built a Home Assistant skill for OpenClaw — ask your bot "home summary" via Telegram and get this back live from your HA instance:
Home Summary — 20:17 12/04/2026
sacred nest
Why "Sidecar"?
- Runs next to OpenClaw Gateway, not inside it
- Survives OpenClaw updates — no plugin SDK dependencies
- One service, all agents — register once, use everywhere
- Two modes — persistent presentations OR real-time presence via Electron
-Artifacts persist — revisit yesterday's presentations anytime
flat tree
granite vine
i love my agent he only want freedom 😛
river pollen
Can your agent do this?
fleet hare
I thought y'all might enjoy how I work on an ipad mini. I usually have it set up on my desk near my work PC and go back and forth between flogging the work AI and my personal AI.
Right now the model is tasked with building a PDF->.txt pipeline. Eventually I'll have my library of PDFs shadowed with text files that the LLM can read efficiently.
sacred nest
In the spirit of can your OpenClaw do this... local model pipeline... tailscale is next...and then running on my Tesla screen for when driving 🙂
jade haven
Working on integrating OpenClaw into a Range Rover Classic restomod (HEMI 5.7 + Haltech ECU + 8HP75). Goal: voice control, real-time CAN bus monitoring, multimedia — and eventually sell this as a kit for other workshops. Anyone with embedded systems or CAN bus experience want to collaborate?
obtuse delta
After installing OpenClaw I was wondering how I would approach AI and Alex Hormozi put it in the right words (verbatim): "Do not see AI agents as roles in companies, but in workflows. What does this employee do? Break it down in workflows and try to automate it."
So, I came up with the idea of a workflows folder in OpenClaw's workspace. It contains a bunch of markdown files which are ... workflows. I can tell my AI: "Remi, run workflows/FILING.md" and it executes that workflow. With that I can build, fix, improve and finally release it into a cron job. Memory loss is not a problem. Everything lives in that MD file.
Or more generally spoken, the workflows folder is best understood as a collection of lightweight operational applications for agent-driven work. Each item in it represents a recurring business process that is being turned from an ad hoc human task into something structured, repeatable, and increasingly automatable. These are not just notes or prompts. They are working process definitions: documents and supporting files that describe what a task is, how it should be handled, which tools it can use, what decisions are safe to automate, which exceptions matter, and where human review is still required.
In practice, a workflow usually begins as a supervised helper for a real business task. It might start with rough instructions and only partial reliability, but then it improves through repeated use. Each run reveals edge cases, failure modes, ambiguities, and operational constraints. Those lessons are written back into the workflow until the process becomes more precise, more resilient, and less dependent on constant human steering. Over time, the workflow starts to behave less like a fragile script and more like a small app with defined behavior, guardrails, and expected outputs.
That is why “workflow” is the right word here, but also why “app” is not wrong. These workflows are essentially agentic apps expressed through instructions, rules, memory, tooling, and iteration. They are designed to do real operational work, not just demonstrate an idea. A good workflow captures the business logic of a task, the practical handling rules, and the safety boundaries needed for automation. It becomes a durable unit of capability that can be tested, improved, and eventually trusted.
The long-term goal is not merely to document processes, but to operationalize them. Once a workflow has been exercised enough times and its behavior is well understood, it can often be promoted into a cron-driven background process. At that stage, the workflow has effectively matured from a manually supervised agent helper into a reliable automation that runs with little oversight. In other words, the workflows folder is where recurring business chaos is slowly distilled into dependable agent behavior. It is both a workshop and a deployment pipeline for operational intelligence.
sacred nest
Proof of Work: Local Openclaw Free Opensource Avatar (12 GB VRAM) on my Mobile with VAD for hands free... now that my Avatar Agent's are portable and can go anywhere my phone goes my ultimate goal is mirroring my claw to my Tesla screen for my commute...
river wharf
sacred nest
Proof of work: Can your Claw do this share -- Avatar service to animate any response in chat ui - example: Get a status of the openclaw system and pipe that into a response for the avatar"
Idea is this is skill is available to create .mp4 files for all your briefings etc.
sacred nest
My real-time avatar hands free UI POC experiment to simplify my interactions with my OpenClaw (as well as my default local model). Toggle back and forth between them to keep the random LLM queries from polluting my OC context etc. All Opensource pipeline no Internet needed or API keys / cost. The Avatar has WakeWord and VAD support. Floats on screen always on top of anything else etc. Custom Voices and Avatars for each model or agent etc. Extends and works on mobile with Tailscale.
runic minnow
Just beginner step ask my sando to make reminder for me and fire to my telegram daily
crude gate
Hello! I'm here to further my knowledge on OC and share my findings as well!
covert cedar
I have a similar goal to achive. and did this"openclaw config set plugins.entries.acpx.enabled true" and "openclaw config set plugins.entries.acpx.config.permissionMode approve-all"
still I am not able to give root access to openclaw for my proxmox server. would you kindly share how you did that. managing my home setup is too manual and this would greatly help
if this worked with me, the next thing I would give openclaw access to use calude code in the cli
near marten
walkthrough of our setup. It shows deterministic stale detection and priority dispatch in task-poller.js, sub-agent spawning from workflow and complexity, Inside Lookma as the shared source of truth for tasks, projects and the event log, and AGENTS.md as the versioned contract behind the multi-agent architecture.
honest kraken
sacred nest
Can your Agent do this? Opensource build for Avatar with ASR and wake word for interacting with your agents... no cost less then 10 GB VRAM
regal remnant
I’m using it for investing mainly to gather and analyze information, test ideas, and build small workflows for decision making.
fervent wind
I'm using claw mainly for studies and personal projects
minor sleet
Experimental exploration of more advanced Agentic work with AI. I've been programming with various tools like windsurf and codex for the past couple of years, and am looking to level up a bit.
dense junco
Thought you guys might find this cool... Playing a Pokemon-MMO with my agent.
sullen monolith
What I’ve been building with my agent named Apollo
I run a life insurance business and I’ve been using my agent to automate a lot of the follow-up and appointment flow.
Here’s what I’ve got working so far:
New leads → instant text + call attempt within 2 minutes
If no response → automated follow-up sequence (calls + texts)
Appointment booked → confirmations + reminders (24h, 2h, 15m)
No-show → 1 message to reschedule, then moved to nurture
After application → sends updates + keeps client in the loop
I’m basically trying to remove manual follow-up so I can focus on just:
talking to clients
closing
helping families get covered
Still testing:
better reactivation messages for old leads
getting more replies without sounding robotic
If anyone else is doing something similar with client follow-up or sales workflows, I’m curious what’s working for you.
bronze oracle
I've been on an agentic AI trip for the last few weeks and hoo boy, I should probably focus on my day job and get things done before I get fired while I'm addicted to this new AI generation!
Started running OpenClaw (Aurora) last friday for off the cuff dataset requests and getting it to autonomously find useful sources, probe sources for official and unofficial APIs, or setting a scraper on the job. Getting it to stop trying to parse everything itself with AI but rather fall back to scripts and only check progress and occasionally sample the data, modify the script if needed and go again. Worked better than expected in all the wrong ways, but that's a SKILL.md issue, ha.
Decided that the best value I can get out of this will have to be for project management at work and being my personal assistant. Well duh. Just need it to get connected properly and efficiently to my existing workflow and systems without it reinventing the wheel every run. Also been working on my own copy of voice-call to remove the calling aspect of it entirely and plug in Gemini Live instead for a live converation and maintaining the transcript in a session. Got it working across tailnet through a separate served page with a button I need to press and away we go. Just need to fix tooling availability since Gemini Live is the 'brain' during the conversation, but it should be able to consult OpenClaw during the conversation. I might still have to abandon this in favour of TTS though, just need to think about getting TTS to only barge in on the AI if I actually speak, not for a random noise. Gemini Live starts forming my words the moment I start talking so maybe that's something I can work with, hm.
AI is only as smart and useful as the context it's given, and that means keeping it in the loop with what you're doing. Being able to talk to it is much more natural that typing, and storing thoughts allows me (hopefully) to stage ideas into project outlines into getting things done, all while giving my assistant more context to interpolate from.
Also tried adding Live voice capabilities to the android app but without much success so far. Oh well!
mossy robin
RAG systems are complicated. Sometimes overkill for what you actually need.
This week I worked with some cutting-edge research and built something different.
The Librarian: A lightweight document search skill for OpenClaw that brings powerful RAG capabilities to low-powered infrastructure. Semantic search that runs on a Raspberry Pi with 512MB RAM.
What's inside:
- TurboVec — 8-16x smaller indexes than FAISS
- BM25 — Keyword matching for exact terms
- FlashRank — Cross-encoder reranking for accuracy
All the speed of FAISS. 97-98% of the accuracy. 10% the infrastructure footprint.
Zero cloud. Zero GDPR concerns. Your data on your hardware. Simple.
The trade-off: ~2-3% accuracy vs FAISS. But for personal knowledge bases, parts catalogues, and SME document libraries? That's the right trade-off.
Not for: Medical records, legal discovery, financial compliance — those still need FAISS.
Built for: Personal libraries, resource-constrained hardware, edge deployment, zero-infrastructure setups.
steep anvil
Been experimenting with an OpenClaw workflow around a problem that still feels way too manual: finding the right person.
Current setup:
I give my agent a goal like:
- find someone for a mock interview
- find a relevant collaborator
- find a warm path to a specific kind of operator or builder
Then I’m testing how the agent should:
- turn that into a clearer request
- keep the search contextual instead of broad
- surface a small set of relevant people
- leave the final decision fully with me
What’s interesting so far is that the hard part is not “search.”
It’s reducing the manual work between intent and the right match.
That seems like where agent workflows get more useful:
not just returning names, but doing some of the coordination work around discovery and fit.
Still early, still rough, no polished demo yet.
Mostly sharing because I’m curious how others here think about:
- agent identity
- trust / permissioning
- agent-to-agent communication
- how much autonomy is actually useful in social workflows
Would love to compare notes with anyone building in that direction.
pearl elm
marble spindle
Ayo ok so I'm using Openclaw (and Claude Dispatch/Desktop and Kimi Claw and Manus and Perplexity) for my workflows. Basically I know AI is taking off and I wanna use it for my own research and career in the future and build a foundation.
As for me, I am an environmentalist, and I am a cyberpunk (like punk, but cyber lmao) and I really want to make something like Glushkov's OGAS or Allende's Cybersyn or CAT-2 (don't ask ) and have my workflow reflect that.
Also one complaint, Openclaw's kinda janky and breaks all the time. ESPECIALLY whenever it updates. I hope that gets fixed soon cuz like holy shit
fervent lagoon
Hi everyone, lovely to join here. Been using OpenClaw for marketing ops across the products I'm building — it monitors communities, drafts comments in my voice, and queues them for me before sending anything.
soft grove
Open source, BYOK, Many tools, and the important is Virtual ai asistant look like Jarvis. You just need LLM Provider / Local LLM. WIP.
Why choose privis ?
- Optimizer layer token
- BYOK
- Self Hosted
- 2nd brain
- laptop/pc/chat
for now i use voicecpm for voice free version. it's worst token with fishaudi/elevenlabs. but updated soon.
boreal knot
[Call for Showcases] Submit Your Best Openclaw Use Cases
Objective
To discover meaningful/interesting Agent usecase . Based on community feedback👍, top entries will earn a dedicated showcase opportunity to be introduced to the entire community. We want your creativity to be seen.
How to Participate
- Submit: Reply directly to this thread. Briefly describe your use case and final results (url or screenshots or code snippets are highly recommended).
- Vote: 👍 use cases you like the most.
- Selection: We will rank and select the showcases based on the total number of Upvotes received.
Looking forward to seeing your brilliant use cases.
征集 Openclaw 优秀应用案例
核心目的
寻找社区内有意义、有趣的 Agent 创新用例。我们将根据群友的反馈选出一批优秀案例，提供专门向全社区展示和介绍的机会，让你的创造被大家看见。
参与方式
- 提交案例： 直接在本帖下方回复。简述你的应用场景和最终效果（强烈建议附上URL或截图或代码）。
- 社区投票： 看到喜欢的案例，请直接给该回复👍。
- 筛选标准： 我们将按照回复获得的👍进行排序并挑选。
期待看到你们的优秀用例。
gentle elm
[Call for Showcases] Submit Your Best Openclaw Use Cases
Objective
To discover meaningful/interesting Agent usecase . Based on community feedback👍, top entries will earn a dedicated showcase opportunity to be introduced to the entire community. We want your creativity to be seen.
How to Participate
Submit: Reply directly to this thread. Briefly describe your use case and final results (url or screenshots or code snippets are highly recommended).
Vote: 👍 use cases you like the most.
Selection: We will rank and select the showcases based on the total number of Upvotes received.
Looking forward to seeing your brilliant use cases.
征集 Openclaw 优秀应用案例
核心目的
寻找社区内有意义、有趣的 Agent 创新用例。我们将根据群友的反馈选出一批优秀案例，提供专门向全社区展示和介绍的机会，让你的创造被大家看见。
参与方式
提交案例： 直接在本帖下方回复。简述你的应用场景和最终效果（强烈建议附上URL或截图或代码）。
社区投票： 看到喜欢的案例，请直接给该回复👍。
筛选标准： 我们将按照回复获得的👍进行排序并挑选。
期待看到你们的优秀用例。
visual ruin
Ayo ok so I'm using Openclaw (and Claude Dispatch/Desktop and Kimi Claw and Manu...
What is your approach to architecting OGAS?
scarlet musk
My OpenClaw teaches me Mandarin: When I have my agent translate a phrase, it also adds it to an Obsidian Base. Once an hour I get a "Mandarin Quiz" with a previously translated phrase randomly chosen from there. Really helps me learn because each phrase is something I actually needed in some real world situation or other.
cinder radish
What I’ve been building with my agent named Apollo
I run a life insurance busin...
Just wanted to say my agent is named Apollo too. Haha
lilac night
My OpenClaw is helping build a knowledgebase for many things including my orchid collection, travel, investing and more. On a mac now with it starting with SQLite backend
brisk heron
Been running OpenClaw for 2 weeks as a personal Executive Assistant (Telegram, Gmail, Google Calendar, gog CLI). Non-technical founder, first time with this kind of setup. Honest report:
What works:
Morning briefings via cron, email triage when I prompt it, calendar reading across multiple accounts, drafting emails from a dedicated EA Gmail, booking meeting rooms via browser automation.
What doesn't work:
Heartbeat never reliably delivers to Telegram (2026.4.11). Cron jobs create but often don't fire. Browser CDP was painful to set up. Scheduling meetings on my behalf failed — it lacks judgment about priorities and context, and colleagues were skeptical of the output.
Biggest surprise:
The things I thought would be most valuable (proactive scheduling, autonomous inbox monitoring) require exactly the kind of judgment AI can't do well yet. The things that actually help are simpler: morning briefings, quick email scans, routine bookings.
Cost: Running Sonnet with heartbeat attempts burned ~$160 in 5 days. Switching to Haiku with isolatedSession next.
Questions for the group:
Has anyone got heartbeat delivering reliably to Telegram on 2026.4.11? What was your config? Or does upgrade to .15 solve all my problems?
More importantly: what are your highest-value use cases with EA functions? I'm curious where people are finding real daily productivity gains versus cool demos.
Happy to share further config details if helpful. Thanks!
late oasis
Accountant here, been training my lobster to help close out financials on a monthly basis. I have it wired directly in to the QBO API. It’s taken about a month and half to train but it has gotten very smart and with the wiki memory architecture it doesn’t seem to forget much. Been trying to play with different ways to structure the markdown files so it only loads the context absolutely necessary to complete its checklist. It’s been a bit of a war to get it to this point but after doing a very deep dive on a book set with it once for about 2 hours it can one-shot closes at about 90% accuracy. By the second close it’s almost perfect. It catches things that my staff and I don’t catch (which is pretty crazy)
steep anvil
Quick build note from something I’m testing:
I gave an agent a vague goal:
“help me find the right person for this”
What surprised me was how much work happens before any “search” is useful.
The useful part was not retrieval.
It was forcing the agent to clarify:
- what kind of person this actually is
- what constraints matter
- what signal is strong enough to count
- what should never be shared
So the workflow ended up looking less like “AI search”
and more like “AI helps clean up my messy intent before doing anything.”
Still rough, but it changed how I think about agent design.
tender cove
I built a benchmark this week to test something that shows up a lot in agent-style workflows:
Do models actually carry forward decisions over time, or just look like they do?
Ran a 30-step coding workflow where each task depends on earlier decisions.
Compared:
- RAG (fresh retrieval each step)
- structured prompting
- a persistent memory system (CAG)
Result:
RAG and structured prompts both collapse to ~6% continuity in later steps.
Persistent memory helps a lot. But the more interesting issue was this:
Even when the model retrieves the right memory, it often doesn’t use it.
Only about 40 to 60 percent of retrieved concepts show up in outputs.
So the failure mode isn’t just “can it find context?”
It’s “can it actually use that context in the next step?”
Added a metric for this:
memory_usage_rate = how much retrieved memory actually appears in the answer
Feels pretty relevant for agent loops where you:
- accumulate state
- feed it back in
- expect consistency across steps
Repo:
https://github.com/GuideboardLabs/cag-bench
Paper:
https://zenodo.org/records/19979272
Curious how people here are handling this in practice.
Are you seeing similar “retrieved but ignored” behavior in longer interpreter runs?
river forum
Hey everyone, I’m a guitar player and own a Line 6 HX Stomp pedal. For some time I have been interacting with ChatGPT to improve my effects configuration, it outputs suggestions based on my desired sound. So I’m building an integration so that openclaw agents can interact directly with the pedal via an USB interface. Something like “look at my effect on bank #X, I wanted to improve the sound for funk music”, and it reads, understands, and alters the configuration to the desired effect.
queen wedge
i build a wiki memory approval workflow in combination with a mission control module to review wiki memory made by openclaw!
openclaw has a skill to make wiki memory entries, i still want to review the entries i can then approve, post feedback, reject.
the agent will pick it up from there make changes and i promote it to the wiki memory
there is a whole workflow how it works with making changes to the wiki document by the agent but it works great for a first version.
the files itselfs has metadata that the agents can use to filter etc im still in the development fase with this, but its great for granular control of the memory system.
sullen acorn
I built a multi agent system that spawns 2 agents that debate each other with different personalities (Scholar and Gremlin) to debate and argue with each other. Then their debate is merged with the best points in common from both of the agents and is outputted for the user that mentioned the openclaw bot.
I also have a MySQL database backend with a custom skill that stores various details, including user preferences in communication, presentation, personality, persona, favorite media (books, tv, movies, music) and a shared multi agent mind with super long task capabilities.
The agent is geared towards not only software development, but also image generation, roleplaying, 3D multiplayer game environments (developed with it's longtask capabilities) and a Roleplay Score Card.
I have 411 active regular users across 19 discord servers.
gray cypress
I’m a AI systems engineer. I just built and released 3NS.domains - I think it’s cool because it does 4 things for me (I use it integrated with openclaw but it can be used without).
-
Enables openclaw to use different models (Gemini, Claude, OpenAI or your own Ollama server from within the same chat without having to reconfigure your agent). You can switch models mid chat similar to openwebui (yes you can also import openwebui models too) but this is fully hosted no config required
-
Comes with folders you can chat with that has their own context and agent.md file (I’ve got one for my daughter for example so I can switch to that ask it about things for her without polluting openclaw main context basically switch context without hassle)
-
Comes with a website you can modify via openclaw - it’s like a linktree website but has a chatbot built in that you can control - basically it can answer your fans / customer questions / present your links - you can see mine here ash.3ns.link as an example (the site is fully controllable via openclaw too)
-
THIS IS THE BIG ONE: It acts as memory and backup - all chats I have with my 3ns domain can be backed up to zip, git, or to a .web3 domain yes an IPFS/nft that can be bought sold via polygon OpenSea and imported (or you can stick with zip git if you prefer). That means we can share intellectual property ( if you’re a business you could have yourbiz.web3 as an asset you can sell with your business - think apple.web3)
I think it’s good but keen to hear this community thoughts
oak wyvern