Forum

Hi,

We checked your server (Aliyun Tokyo 4C 8GB) and found it's in an unhealthy state:

• SSH port open but connections fail during banner exchange
• K3s API port open but connections get refused or time out

This means the VM is running but both SSH and K3s are overloaded or unresponsive — which is why OpenClaw can't start.

Please try:

1. Reboot the server — Dashboard → Servers → Aliyun Tokyo 4C 8GB → Reboot
2. After reboot, wait 3-5 minutes for K3s to fully initialize
3. Then try starting OpenClaw again

If the server remains unresponsive after rebooting (we've seen several recent reports of Aliyun VPS instability), we'd recommend:

• Deleting this server (refund goes to account balance automatically)
• Purchasing a server from a different provider like Tencent Tokyo, Lightsail Tokyo, or Hetzner — these have been more stable recently

Let us know how it goes.

Thanks for your reply Bohan, I've just purchased a new server (Hetzner) with same specs (4C, 8GB RAM), will see if this works or not.
In the meantime, I will delete the Aliyun server & request for refund.

Sounds good! Hetzner has been reliable in our experience. After you delete the Aliyun server, the refund will be automatically applied to your account balance.

Let us know if you run into any issues setting up OpenClaw on the new Hetzner server.

I still cannot start the OpenClaw using Hetzner server. Zeabur AI's tried to solve different issues but "502 Bad Gateway Error" kept repeating. The AI has sent a report to Zeabur technical team for me. Please help, thanks.

Hi dom,

We checked your OpenClaw service and it's currently RUNNING normally on the Hetzner server.

One thing to note: OpenClaw takes longer to initialize compared to typical services. Based on the logs, the startup process takes roughly 1.5–2 minutes when dependencies are already cached — and potentially 3–5 minutes or more on a fresh install, as it needs to set up Homebrew, Go, and the gateway before it's ready to accept connections. During this window, you may see a 502 error, which is expected behavior.

Could you try accessing domclaw333.zeabur.app now and let us know if everything is working as expected?

Hi Can,

Thanks for your prompt reply.

I waited one hour after starting my OpenClaw. The control UI showed "RUNNING" but the Web UI showed 502 error. The "龍蝦面板" was not available all the time (the panel kept loading all the time). I'm sure that there are some problems with my OpenClaw.

I deployed the OpenClaw using Zeabur template & could never start the service so far. Zeabur AIs tried to fix different issues for me but didn't help. Please do look into my case & help resolve all the problems. I appreciate your help !

Hi dom,

We've investigated your case thoroughly and resolved the issue. Here's a full explanation of what happened.

Root Cause 1: Out of Memory (OOM)

Your Hetzner 4C 8GB server ran out of memory while running OpenClaw. The service consumed up to ~6 GB of RAM during peak usage, causing the node to run low on memory and triggering automatic pod evictions. This caused OpenClaw to restart repeatedly throughout today (5+ restarts in total).

We've noticed that OpenClaw tends to consume all available resources on the server if no limits are set. To prevent this from destabilizing your server, we've set resource limits on your service: CPU 2500m and Memory 4500 Mi. Please note that if OpenClaw exceeds the memory limit during heavy usage, it will restart automatically — if you notice frequent restarts, this is likely the cause.

Suspected Cause 2: Routing misconfiguration

After the repeated restarts, we believe the internal network routing (Ingress) became out of sync with the actual port your service was listening on — which would explain why traffic was being directed to the wrong port and returning 502 even when the Dashboard showed "RUNNING". We're still investigating the exact root cause of this. In the meantime, we've manually corrected the routing configuration, and the service is now fully accessible.

We also noticed the network port was previously changed to 18791. Could you let us know when or why that was changed? We've reset it back to the default value of 18789 for now.

Current status:

• domclaw333.zeabur.app is returning 200 and loading normally ✅
• Memory usage is currently ~656 Mi (healthy, at idle)

Sorry for the inconvenience — please let us know if you have any further issues!

Thanks for your reply Can. Unfortunately, the problems persist.

The system ran properly after copying and pasting the gateway token at the beginning. However, the "龍蝦面板" was still not available.

Somehow it was disconnected (1006) : no reason & then 502 error appeared again after a few minutes.

I am not familiar with coding at all. I just deployed the OpenClaw with Zeabur template without any modifications. Please help find out the root cause, thanks.

Hi Dom,

We've identified the root cause of your repeated crashes. A background process (openclaw models list --all --json) is hanging indefinitely with no timeout, causing memory to grow until the pod hits the 4500 Mi limit and gets OOMKilled. This triggers the WebSocket 1006 disconnects and 502 errors you're seeing.

Temporary ask: Please stop using the OpenClaw panel (龍蝦面板) for now, as interacting with it triggers the problematic model enumeration process.

For configuration, please refer to the official OpenClaw template documentation: https://zeabur.com/templates/VTZ4FX

This covers the recommended setup and environment variable configuration. Once you've set things up according to the template docs, the service should stabilize.

We're also looking into better solutions for this issue on our end.

Hi Can,

It is said that no setup is required if I deploy the OpenClaw with Zeabur template. Coding knowledge is not required. Do not understand what configuration I need to set now.

Maybe redeploying a new OpenClaw is a quick & easy solution for me. Please advise.

I never used the OpenClaw panel as it is not ready to use so far. I feel frustrated and tired after fixing the system back and forth. Could you please help me out ? Thanks.

Hi dom,

I hear you, and I want to be honest about the situation rather than keep chasing symptoms.

The real issue

OpenClaw is a very large, feature-rich service — it bundles an AI agent runtime, gateway, control UI, device pairing, Telegram/WhatsApp integrations, model management, and more. Because of that surface area, "deploy from template and it just works" isn't fully realistic. There are setup steps (Gateway Token, device pairing, and often tuning for which features you actually want enabled) that can't be skipped regardless of the platform.

The Zeabur template gets you a running instance in one click, but the configuration afterward is something you do need to engage with. Redeploying fresh won't remove that step — you'd land in the same place.

What would help most

Instead of going in circles, could you tell me:

1. What do you actually want to use OpenClaw for? (e.g. chat interface only, Telegram bot, WhatsApp, a specific workflow)
2. Which part is blocking you right now? (the 龍蝦面板 loading, connecting the Control UI, or something else)

With those two answers, I can walk you through exactly the steps you need — and skip everything that isn't relevant to your use case. That's likely faster than another full redeploy.

If OpenClaw's complexity ends up being more than you need, I can also suggest lighter alternatives depending on what you're trying to accomplish.

Hi Can,

Am I missing out on something in our previous communications ?

First step to start my OpenClaw. Second step to set the configuration. I am now stuck on the first step & can't set my own configuration like setting my preferences, AI models, skill set, channel, etc.

My OpenClaw ran properly at the beginning whenever I started the service. However, the system would be disconnected (1006): no reason after a few minutes. My AI said that there were still gateway token issues. The token was lost everytime. Something needs to be fixed, Can...

By the way, I could chat with my AI on the 'Gateway dashboard' when the system was connected. However, I couldn't use the 'OpenClaw panel'. The 'Terminal' (under the 'Command') did not respond. Is this normal ?

I do not understand coding at all. I want to use my OpenClaw to run a small business for me. It can do research, trading workflows, simple coding as well as community operations & personal tracking, etc.

I know you are trying to help me out. Really appreciate it !

Hi dom,

Good news first — your OpenClaw is currently running normally. The service is up and the agent is responding. So the deployment itself is fine; what's left is just picking how you want to interact with it.

The best approach: connect a messenger

The simplest and most stable way to use OpenClaw is to connect it to a messenger such as Telegram or Discord. Configure the bot once, and from then on you just chat with your agent from your phone or desktop messenger. This is what we'd recommend for your use case.

OpenClaw supports many messengers — if Telegram or Discord aren't what you want, you can see the full list of supported channels here:

https://docs.openclaw.ai/channels

All connection methods require a pairing step

For security, every client — Web Control UI, Telegram, Discord, or any other channel — needs to be paired once with a device approval before it can talk to your agent. This is by OpenClaw's design, not a bug. Until pairing is done, the client will keep asking for the token / keep getting rejected. Once paired, that client works normally without re-authenticating every session.

The full steps for each channel, including pairing, are documented on the OpenClaw template page:

https://zeabur.com/templates/VTZ4FX

Next step

Please follow the template guide for your chosen messenger (including the pairing step). If you hit a specific step where something doesn't behave as the guide says, send a screenshot of the page you're on plus a short description of what's not working, and I'll take it from there.

Hi Can,

The problems are not resolved. The Gateway dashboard was ready after I copied & pasted the gateway token. However, the system was disconnected after a few minutes. The dashboard was resumed & then disconnected again. It kept looping back around.

My agent said that the system couldn't keep the token persistently. So I let it fix it.

At the end, it came up with the following conclusion :

Root cause :

- Token environment variable not persisting across restarts

- Read-only filesystem blocking script modification

Solutions :

- Set OPENCLAW_GATEWAY_TOKEN as a container/environment variable

- Create an entrypoint script that loads the token file

My agent can't modify start_gateway.sh & change container configuration.

Can, please review if you agree to the above findings & help make all necessary corrections. Hopefully, all the problems can be resolved once and for all, thanks.

Hi dom,

We reviewed the AI agent's findings and your logs. Here's a summary:

The server-side setup is fine. The Gateway Token is correctly saved in your service's environment variables and does not get lost when the service restarts. The AI agent's diagnosis about "token not persisting" was incorrect — no fix is needed on the server side.

Why you keep having to re-paste the token: The token you paste in the Gateway dashboard is only remembered by your browser for that session. Once the connection drops (e.g. due to idle timeout or closing the browser tab), your browser forgets the token, and the next time you visit the page you need to paste it again. This is how OpenClaw's security works — it's not a bug on Zeabur's side.

Our recommendation:

Instead of using the Gateway dashboard in your browser, we recommend connecting OpenClaw to Telegram or Discord. This is a one-time setup — once connected, your AI agent will always be available through the messenger app, and you won't need to deal with token pasting or disconnections.

Which messenger app do you normally use? We can walk you through the setup step by step.

Hi Can,

Understand your point. However, the system was disconnected itself everytime & I didn't close the browser nor exit the service. It always showed "error" (see the screenshots attached - "error1" & "error2"). The Gateway dashboard was not available & I couldn't chat with my AI agent whenever the system was disconnected. So how can I set my configuration ?

There is always a small window on the 'Zeabur dashboard' showing a 'Gateway dashboard' (gateway token is required) eventhough my AI agent is responding (see the screenshot attached). Is that normal ?

The 'OpenClaw panel' is never available for use (see the screenshot attached). Is that normal ?

The Terminal is never responding (see the screenshot attached). Is that normal ?

Hi dom,

Thanks for the screenshots — they helped a lot. We've done a deep investigation into your service and here's a full update.

1. Your gateway is currently running normally

We checked the logs and confirmed that after the brief outage you saw (the "OpenClaw Unavailable" page), the gateway restarted and has been running fine since. The outage was caused by the gateway process exiting on its own (exit code 0, meaning a clean shutdown — not a crash or memory issue). This may have been triggered by the AI agent's earlier attempts to modify system files.

2. The AI agent's diagnosis was incorrect

We verified that OPENCLAW_GATEWAY_TOKEN is correctly set in your environment variables and persists across restarts. There is no need to modify start_gateway.sh or create an entrypoint script — the token configuration is working as intended.

3. About the small Gateway Dashboard window on the Zeabur Dashboard

The small popup you circled in red is a service preview widget built into the Zeabur Dashboard. It always shows the login page because it doesn't carry your token. This is normal — it's not an error. To actually use the Gateway Dashboard, open domclaw333.zeabur.app directly in a separate browser tab and paste your token there.

4. Do NOT use the OpenClaw Panel tab

The OpenClaw Panel (龍蝦面板) is an experimental feature and is the cause of your crashes. When accessed, it triggers background operations (such as model enumeration) that consume too much memory and cause the gateway to crash. Please avoid using this tab until it's stable.

5. Most important next step: connect a messenger

The Gateway Dashboard in the browser is only meant for quick access and initial setup. The way OpenClaw is designed to be used day-to-day is through a messenger like Telegram, Discord, WhatsApp, etc. Once connected, your AI agent is always available through the app — no token pasting, no disconnection issues, and you can chat with your agent anytime from your phone or desktop.

Please follow the step-by-step instructions on the template page to connect your preferred messenger:

https://zeabur.com/templates/VTZ4FX

Which messenger would you like to use? We'll walk you through the setup if you need help.

Hi Can,

The problems persist. I am so frustrated. The gateway dashboard was disconnected (1006) : no reason after I started the service & chatted with my agent for only 2 minutes. Then the attached error message popped up again & I coudn't restart the service.

I didn't open the 'OpenClaw panel'. I didn't open the 'Terminal' & I didn't close the browser. No input from My OpenClaw so far but I am paying for the platform & my AI agent almost everyday...

Hi dom,

We've been looking into the disconnection issue and suspect it may be related to the OpenClaw version your service was running (2026.4.5). Several stability fixes have been released since then, including:

• Improved gateway restart handling and connection resilience
• Better memory management for background processes
• Fixes for WebSocket session state preservation

We've updated your OpenClaw to the latest stable version (2026.4.26). The service has been restarted and is currently running.

Please try using the Gateway Dashboard again and let us know if the disconnections still occur.

One thing to note: OpenClaw is under very active development, which means occasional instability between versions is expected. If everything works well on this version, we'll keep it as-is. If the issue persists, we'll continue investigating from there.

Hi Can,

It seems okay now. The system was disconnected once but re-connected automatically in a few moments. No disconnections in the next 60 minutes. The 'Terminal' is also responding now.

However, my AI agent's responses were extremely slow (the previous agent responded more quickly). The operations on the Gateway dashboard were also slow. It was hardly to scroll up/down the screen under 'Communications/Channel' (extremely slow reaction). It looked like insufficient RAM or the CPU was running slow. Are there any issues with the system ? Could you fix it ?

I would like to set my channel to Telegram. Can't find the setup in detail on the OpenClaw guidelines. It would be appreciated if you could help.

Hi dom,

Glad to hear the disconnection issue is resolved with the version update!

Why it was slow

We checked the metrics and found the cause: Node.js was configured with a 2 GB heap limit, but your OpenClaw was using ~1.97 GB — nearly at the ceiling. When memory gets that close to the limit, Node.js spends a lot of time on garbage collection instead of processing your requests, which is why everything felt sluggish.

What we've changed

Since your server has 8 GB of RAM and OpenClaw is the only service running on it, we've increased the resource allocation:

• Container memory limit: 4.5 GB → 6 GB
• Node.js heap limit (NODE_OPTIONS): 2 GB → 5 GB

Please restart the service from Dashboard → Service → Restart to apply these changes. You should notice a significant improvement in response speed and dashboard responsiveness afterward.

About the Gateway Dashboard

The Gateway Dashboard (the browser UI at domclaw333.zeabur.app) is designed for initial setup and configuration only — it's not meant for day-to-day use. For regular conversations with your AI agent, we recommend connecting a messenger like Telegram, which is much more stable and convenient.

Telegram setup

Please follow the step-by-step guide on the template page:

https://zeabur.com/templates/VTZ4FX#telegram-bot-setup

If you get stuck at any specific step, send a screenshot and we'll help you through it.

Let us know if the speed is better now!

Hi Can,

No change. Both AI agent & the dashboard are extremely slow. I did restart my service : Zeabur dashboard -> Setup -> general setup -> restart all services.

Please help, thanks.

By the way, I've successfully linked to Telegram.

Hi dom,

Great to hear Telegram is connected!

We found the cause of the slowness. Your AI agent has a built-in browser tool (for web research, etc.), but Chrome couldn't launch due to a container security restriction. Every attempt timed out after 60+ seconds, causing your sessions to freeze for 2–3 minutes at a time — which is why both the agent responses and dashboard felt extremely slow.

We've fixed the Chrome configuration on your service. Please restart the service (Dashboard → your OpenClaw service → Restart) to apply the change.

After restarting, you should notice a significant improvement in both agent response speed and dashboard responsiveness.

One thing to note: Now that the browser tool is working, it will use additional memory when your agent browses the web. If you notice the service restarting on its own during heavy use, it may be running out of memory. Let us know if that happens and we can adjust the resource allocation.

Hi Can,

Thanks for your help. However, I tried & both AI agent & the dashboard were still extremely slow right after you reallocated the memory. Chrome was not installed at that time.

Chrome was installed today as we tried an alternative for my agent to check gmails. Anyway, we didn't use it as we know Chrome doesn't work for AI agents. Chrome is not the cause of slowness. Please help.

I've found the following problems today :

I can see that there are some skills pre-installed & I've also installed some more like 'Talk' & 'Voice' skills (please see the photo attached). However, my agent can't "talk". I did restart the service after installation. Do I need to do further steps to implement the skills. So my agent can really use them ?

Some skills are said to be "available" for installation but it said "Rate limit exceeded". What does it mean ? Please see another photo attached.

Big files cannot be transmitted through the dashboard nor Telegram. There are 150GB space on my Hetzner server. So my agent can install & save apps. Is it possible for me to upload big data files & even install some apps onto my server for agent to use ?

I failed to send a screenshot (jpg) to my agent through dashboard & Telegram. At the end, my bot was disconnected from Telegram & showed "405 litellm.APIError: APIError: DeepinfraException - Error code: 405 - {'detail': 'Multimodal is not supported for model: {inp.model}'}. Received Model Group=glm-4.7-flash. Available Model Group Fallbacks=None". I can't chat with my agent with Telegram now. Please advise.

Thanks a lot.

Hi Can,

Problems with Gateway dashboard & Telegram, please help !

My agent attempted to use the skills installed in the OpenClaw yesterday. Then Telegram kept showing : "Something went wrong while processing your request. Please try again, or use /new to start a fresh session" & cannot be restarted now.

I checked that Telegram token was still kept in the system. So I pasted the Telegram pairing code again in the Terminal. The service cannot be restarted & the Terminal is not responding somehow. I can't communicate with my agent on dashboard nor Telegram now. Please help !

By the way, updates are always available for OpenClaw. Shall I keep updating it by pressing the top right "update" button on the gateway dashboard ? Thanks.

Hi dom,

We checked your service and it looks like things are back to normal after your recent AI Hub top-up — your agent is running and responding on Telegram successfully.

Here's what happened: your AI Hub balance had been fully consumed, which caused every AI model request to fail. That's why neither the Gateway dashboard nor Telegram could get a response from your agent. The top-up resolved this.

A few tips to avoid this in the future:

• Enable Auto-Recharge in Zeabur Dashboard → AI Hub, so your agent doesn't go silent when the balance runs out.
• You can also use your own AI provider API key (OpenAI, Google Gemini, Anthropic, DeepSeek, etc.) instead of AI Hub. This gives you more control over model selection and billing. You can configure this in Gateway dashboard → Settings → Model Provider.

Answering your other questions:

1. Screenshot breaking Telegram: Your previous default model (glm-4.7-flash) did not support images. To send images to your agent, make sure you're using a multimodal model (such as GPT-4o, Gemini, or Claude Sonnet). You can check and change the default model in Gateway dashboard → Settings.

2. Skills (Talk, Voice, etc.): Some skills require specific model capabilities or external services to function properly. The "Rate limit exceeded" error during installation is a temporary throttle from the skill marketplace — wait a few minutes and retry.

3. Big files: The Gateway dashboard and Telegram have file size limits. For large files, upload them to a cloud storage service and share the link with your agent.

4. "Update" button on the Gateway dashboard: This updates OpenClaw by changing the image tag in your service settings. You can also change the image tag yourself in Zeabur Dashboard → Service → Settings if needed. However, we recommend staying on the current version (2026.4.26) for now, as it resolved your earlier disconnection issues. Only update if you experience specific problems that require a newer version.

Let us know if you have any other questions about configuring your agent!

Hi Can,

Thanks for your reply.

I've changed my AI agent from glm-4.7-flash to Gemini 2.5 flash lite in order to complete different tasks including image editing, etc with faster responses. However, there are quite many problems when communicating with my agent in Telegram after AI Model Switching. My agent can no longer hear voice messages, forgets how to communicate using my preferred language (writing traditional Chinese & speaking Hong Kong Cantonese), has difficulty searching info (extremely slow), etc.

So I decide to update my OpenClaw. I can't see any option to update the OpenClaw in Zeabur Dashboard → Service → Settings but there is always a button on top of the Web UI (Gateway dashboard). I'm using the service under Zeabur environment instead of my local PC. Believing that most of the users in Zeabur will experience the same problems/bugs, if there are any, after the update.

Again, it seems that my agent does not know how to use the skills installed in the OpenClaw even though same commands are activated in Telegram (no matter using glm-4.7-flash or Gemini 2.5 flash lite). Did I miss anything in the settings of OpenClaw/Telegram ?

Last question, I'm using ZEABUR_AI_HUB_API_KEY. My agent may delegate tasks to other AI models within the Zeabur AI hub. Any differences or advantages if I add subagents to my OpenClaw ? Thanks.

Hi dom,

Thanks for the detailed update. Let me address each point:

1. Why Gemini 2.5 Flash Lite is causing problems

AI models are not interchangeable — each has a different mix of capability, speed, cost, and supported input types (text / image / audio). Quick guide for OpenClaw on AI Hub:

What you're seeing — forgetting language preferences, not invoking skills, slow tool use, voice broken — is Flash Lite being underpowered for OpenClaw. It's built for high-volume simple tasks, not autonomous agents.

Recommendation: switch to claude-sonnet-4-6 — best fit for your use case (research, trading, multilingual incl. Cantonese, voice, images). Costs more per call than Flash Lite but burns far fewer tokens because it actually completes tasks. claude-haiku-4-5 is a cheaper middle ground.

Switch in Gateway Dashboard → Settings → Model, or tell your agent: /model zeabur-ai/claude-sonnet-4-6.

2. Voice messages

Voice broke because Gemini 2.5 Flash Lite doesn't accept audio input (text + images only). glm-4.7-flash could process audio directly, which is why it used to work. Two fixes:

• Switch to an audio-capable model (claude-sonnet-4-6, gemini-2.5-pro)
• Add a separate Whisper audio node — model-independent, more robust: https://docs.openclaw.ai/nodes/audio

3. Updating OpenClaw

The "Update" button on the Gateway Dashboard and changing the Image Tag in Zeabur Dashboard → Service → Settings → Image Source do the same thing.

You're on 2026.4.26. The latest stable is 2026.5.4 — feel free to update. Pin to that specific tag, don't use latest (auto-updates on restart, can break things). Back up your Config from Gateway Dashboard's Config page first, in case fields are deprecated.

4. Skills not being invoked

Reasons skills may not work:

• Model capability — small models often skip tool/skill calls or invoke them wrong
• Some skills need companion services:
• coding-agent → needs OpenClaw DevBox in the same project
• Browser/web skills → need the Sandbox Browser template
• Voice/Talk skills → need a Whisper audio node
• "Rate limit exceeded" on the marketplace is a temporary throttle — wait a few minutes and retry

Installing the skill alone isn't enough; the supporting service has to be there too.

5. Subagents vs AI Hub

Different concepts:

• AI Hub = model router. One API key, many LLM providers. Only decides which LLM answers.
• Subagents = specialized agent personas with their own prompts/tools/model. The main agent delegates tasks to them (e.g. a "researcher" with web tools).

For your use case, you can get most of the way without subagents by:

1. Using claude-sonnet-4-6 as default
2. Writing clear instructions in OpenClaw's memory.md (language, role, tasks)
3. Installing only the skills you need, plus their supporting services

Subagents add complexity — worth exploring once the basics are stable.

Let us know which model you'd like to switch to and we'll verify it's working.

This post has been inactive for a while. We will be closing it in 2 days if there is no new activity.

Hi Can,

Thanks for your info which is very helpful. I've changed to use Claude Haiku 4.5 which offers a good cost-performance ratio.

My AI agent could "hear" & understand Cantonese somewhat using Claude Haiku 4.5 after installing Whisper audio node but still could not speak Cantonese (what else can I do ?) . However, it lost the Whisper audio node somehow while trying to speak. In the meantime, I'm trying hard to explore different approaches to give my agent access to Gmail. Believing that Google Workspace CLI may work for the skill 'gog'. Please comment.

Things are getting back on track after the disconnection issues were resolved but there is still a long way to go before I can expect substantial outputs from my agent. People says that AI tools are accessible to non-technical users. However, plug-in softwares & companion services are always required. Basic coding knowledge is essential. It is not that easy for me to make my OpenClaw to run a business model. It would be greatly appreciated if Zeabur team could help me out.

Hi dom,

Thanks for the update. Glad to hear the disconnection issue is now mostly resolved and that you have switched to Claude Haiku 4.5.

A few clarifications:

1. Cantonese voice

Whisper is for speech-to-text only, so it helps your agent understand voice messages. It does not provide voice output by itself. If you want the agent to speak back in Cantonese, you will need a separate text-to-speech / voice output integration that supports Cantonese.

2. Gmail access / gog

For Gmail access, gog may be relevant:

https://gogcli.sh/

That said, Gmail integration usually involves Google account authorization, OAuth permissions, and security decisions. Please review the gog documentation carefully and start with the minimum permissions needed.

This is application-level configuration inside OpenClaw / third-party tools, so it is outside the scope of direct Zeabur forum support. We can help if the Zeabur service itself fails to deploy, start, route traffic, or persist environment variables, but we cannot directly configure each user’s Gmail workflow or business automation setup through forum support.

3. OpenClaw learning curve

You’re right that many AI tools are advertised as accessible to non-technical users, but the actual learning curve depends heavily on the product.

For example, ChatGPT is designed as a hosted chat product, so the basic experience is intentionally simple: open the app, type a message, and get a response.

OpenClaw is different. It is a highly customizable agent runtime with models, skills, channels, companion services, memory, tools, and third-party integrations. That flexibility is powerful, but it also means the setup is much closer to configuring an automation system than using a simple chatbot. So yes, the learning curve will be much higher, especially when connecting services like Telegram, Gmail, browser tools, audio nodes, or custom skills.

At this point, the Zeabur-side issues we identified earlier have been addressed: the service is running, Telegram is connected, and model/audio behavior is mainly related to OpenClaw configuration, selected skills, model capability, and third-party integrations.

Please note that Zeabur forum support can help with Zeabur platform issues such as deployment, networking, resource limits, restarts, and service availability. Direct hands-on configuration of OpenClaw workflows, Gmail integrations, custom skills, or business automation logic is only available under higher-tier enterprise support.

For regular usage, users are expected to follow OpenClaw and third-party tool documentation and manage application-level configuration themselves.

Since your base setup is already in place — Telegram is connected, Claude Haiku 4.5 is selected, and Whisper is installed for voice input — the next step is not to keep changing the base setup.

Instead, please add only one new integration or workflow at a time, verify it is stable, and then move to the next one. For example, if Gmail is your next priority, focus only on Gmail / gog first before adding voice output, browser tools, more skills, or subagents.

If you need further help with a specific new issue, please open a separate ticket with the exact problem, screenshots, and error messages. That will help us keep each issue focused and route it to the right support scope.

Hi Can,

Thanks for your reply with detailed guidelines for gog. I'll go through all the steps & hope that my OpenClaw can be integrated with Google Workspace soon.

I'll let my agent learn to use skills one by one. Please give advice on the following issues relating to setup in Telegram/OpenClaw :

• The warning "Context limit exceeded. I've reset our conversation to start fresh - please try again. To prevent this, increase your compaction buffer by setting agents.defaults.compaction.reserveTokensFloor to 20000 or higher in your config" will pop up whenever the conversations in Telegram are too long. Zeabur AI tried to help modify the configuration but failed. Could you please help ?
• My AI agent can hear & speak Cantonese now. However, it fails to read MEMORY.md and/or forgets how to pick up the skills again after restart or session renewed although the entire workflow & clear instructions are documented onto the MEMORY.md. Please advise how to make sure that the agent will keep using all the skills automatically every time after restart & session renewed.
• I haven't yet updated my OpenClaw. What version do you suggest now ? Thanks.

This post has been inactive for a while. We will be closing it in 2 days if there is no new activity.

Hi Can,

Please reply to my message above, thanks.

Hi dom,

Sorry for the delayed reply.

Using your AI agent for configuration

All three questions you asked — compaction config, MEMORY.md structure, and skills setup — are things your AI agent can handle directly. Since you're running OpenClaw with Claude Haiku 4.5, your agent has access to the Terminal, config files, and system settings. You can ask it to make these changes for you in natural language.

For example, try telling your agent:

• "Set the compaction reserve tokens floor to 20000 in the config file and commit it"
• "Review my MEMORY.md and restructure it — keep the most important instructions at the top, make it concise"
• "Check which skills are installed and make sure they are enabled in the active config"

If the agent can't complete a task or gives an error, send us the exact error message and we'll investigate whether it's a platform-level issue.

Version recommendation

Since your service has been running smoothly on 2026.4.26, there's no need to update. Only consider upgrading if there's a specific new feature you want that requires a newer version.

Audio issue

We noticed audio transcription has been failing in the logs since around May 19. If you're still using voice messages, ask your agent: "Check and fix the Whisper audio node configuration."

Going forward, for OpenClaw application-level configuration (models, skills, memory, channels, workflows), please try asking your AI agent first — it's designed to handle these tasks. If the agent fails or you hit a Zeabur platform issue (service won't start, networking, resource limits), we're here to help.

This post has been inactive for a while. We will be closing it in 2 days if there is no new activity.

Hi Can,

Many thanks for your reply.

I'd like to know if my OpenClaw can be transferred to a new server or my local computer ?

This ticket has been inactive for a while. We will be closing it in 2 days if there is no new activity.

Hi Can,

Please advise, thanks.

Hi dom,

Yes — your data (memory.md, config, skills) is stored inside your service's container filesystem and is fully under your control. To move to a new server or run locally, ask your agent to export the config files first, then restore them on the new instance. Environment variables can be copied from Dashboard → Service → Variables.

This ticket has been inactive for a while. We will be closing it in 2 days if there is no new activity.