Managing 10 OpenClaw Instances: Multi-Platform Deployment & Real Workflows

My Windows machine has 64GB of RAM. I set it up and haven't touched it in two weeks. It's running 10 OpenClaw instances -- Instance A and Instance B fix each other, no need for me to remote in at all. I also set up a Mac Mini, which the team uses.

People in the community keep asking: why run multiple instances? What do you do when config files keep breaking? How do you auto-post to X? I've compiled all these questions and put together the answers, along with the real workflows I've been running daily for the past two weeks.

Frequently Asked Questions

Q: Why run multiple instances? Can't one bot handle everything?

Two reasons. First, one bot isn't enough -- there are tons of tasks you can hand off to them. Second, and more importantly, you need at least two instances as mutual backups. If one's config file gets corrupted, the other one can fix it.

Q: How much resources do multiple instances need?

Roughly 500MB of RAM per instance. My 10 instances consume about 6-7GB total. The longer you run them, the more local data accumulates -- mine is already at 1GB.

Q: What do I do when model configuration keeps failing?

Avoid editing config files manually -- it's easy to mess up. Two simple approaches:

First, install Claude Code or Codex locally, then have it configure OpenClaw for you.

Second, start with a domestic model. Buy a Zhipu or MiniMax package for a few dozen yuan, then just keep hitting Yes, give it whatever API keys it needs, and get it running first. Once it's up, feed it the aigocode tutorial and let it switch models for you.

The key lesson: set up multiple instances with multiple models, and you'll rarely have issues.

Q: What if the AI doesn't write in my style?

Six months ago, AI wrote worse than me. Six months later, I can't keep up with it anymore. Pick the best model, and I just handle the Yes and No decisions.

Choosing Between Three Platforms

Feishu (Lark)

I started with Feishu because that's where the team does daily work. Just @ in a group chat and you're talking to the bot -- very convenient. But after using it for a while, I found two problems: the chat window doesn't support Markdown output, and streaming responses aren't great.

It's excellent for notifications though. We've connected multiple webhooks, and the Feishu group sends us all kinds of alerts in real time.

The good news is that OpenClaw is building a native Feishu plugin, currently in beta testing. The new plugin supports streaming card replies, emoji reactions, reading merged forwarded messages, and can even operate Feishu's spreadsheets, calendar, tasks, and docs as the user. The previous shortcomings are basically all being addressed.

Discord

Going global means moving to Discord. It felt right immediately. Create a Server, organize Channels underneath: weekly, project, passage, construction. Each Channel can spawn Threads for more granularity -- the main channel stays clean while details fold into Threads.

I recommend Zhixian's article "Agent Trainer Advanced Guide: Building an Efficient OpenClaw Collaboration System with Discord" -- my Discord workspace was built following his approach.

Say one thing in a channel, and the Agent automatically opens a Thread to reply without polluting the main feed. Different Channels can bind different Agents with different models. OpenClaw supports 4 concurrent Sessions by default, meaning you can push 4 threads simultaneously.

Telegram (Primary)

Telegram is lighter, easier to configure, has streaming output, emoji responses, and a great mobile experience -- you can chat with your Agent while walking down the street. Telegram is now my primary platform.

Discord is heavier but has stronger organizational capabilities, making it better for team collaboration.

How to Run Multiple Instances

The core idea is simple: each instance is an independent OpenClaw process with its own config file and Bot Token.

Telegram

If you already have one instance running, spinning up another is straightforward:

1. Find @BotFather and create a new Bot to get a Token

   

   

2. Send the Token to your existing OpenClaw instance and let it handle the rest of the configuration and startup

It will automatically create the new instance's config file and start the new process. Each Telegram Bot is an independent Agent -- chatting with different Bots means talking to different Agents.

Discord

Same as Telegram -- each instance corresponds to an independent Discord Bot.

1. Open discord.com/developers/applications in your browser, click New Application to create a new app

2. Click Bot on the left, click Reset Token to generate a Token, and save it. Then turn on all three Intents below, and remember to click Save Changes

3. Click OAuth2 on the left, then URL Generator. Under SCOPES, check bot and applications.commands. Under BOT PERMISSIONS, check Send Messages, Read Message History, and View Channels. Copy the generated link at the bottom, open it in your browser, and select your Server to authorize

4. Send the Token to your existing OpenClaw instance and let it handle the rest of the configuration and startup

Once done, you'll see the Bot in Discord and can start chatting with it directly.

Each Bot can be assigned to different Channels, each doing their own thing. Bots in the same Server automatically appear in any new Channel you create -- no need to re-invite. But if you create a new Server, you'll need to invite the Bot again.

Discord has another advantage: combined with Threads, the main channel stays clean while details fold into Threads. Different Channels can give different Bots different systemPrompts -- for example, the #weekly channel gets a weekly report assistant, #code-review gets a code reviewer, each handling their own domain.

Feishu (Lark)

Create one Feishu app per Agent. Each app is an independent Feishu bot.

1. Log in to the Feishu Open Platform (open.feishu.cn) and create an enterprise custom app

2. Go to the app details page and get the App ID and App Secret from "Credentials & Basic Info"

3. First go to "Add App Capabilities" and add a bot -- some permissions require the bot capability to be enabled first

4. Go to "Permission Management" and add these permissions: contact:user.base:readonly, im:message, im:message.p2p_msg:readonly, im:message.group_at_msg:readonly, im:message:send_as_bot, im:resource. If you want the bot to respond to group messages without being @-mentioned, also add im:message.group_msg

5. Go to "Events & Callbacks", set the event configuration method to "Receive events via long connection", and add these events: im.message.receive_v1 (required), im.message.message_read_v1, im.chat.member.bot.added_v1, im.chat.member.bot.deleted_v1

6. Click "Version Management & Publishing", create a version, and submit for publishing

7. Send the App ID and App Secret to your OpenClaw instance and let it handle the configuration. Or manually run:

openclaw config set channels.feishu.appId "cli_xxxxx"
openclaw config set channels.feishu.appSecret "your_app_secret"
openclaw config set channels.feishu.enabled true

After publishing, you can chat with the bot in Feishu -- just pull it into a group and @ it.

One gotcha: group chats default to whitelist mode, so the bot might not respond after being added to a group. Just have your OpenClaw instance update the group chat policy configuration.

Feishu uses WebSocket long connections -- no callback URL or public IP needed.

Practical Use Cases

WeChat Official Account Publishing

Before: picking a topic (1 hour) + writing (1 hour) + formatting (30 min) + finding images (20 min) = nearly 3 hours per article.

Now I just tell the AI "write an article about xxx and push it to the WeChat drafts." It automatically gets the access_token, uploads images to the WeChat media library, generates HTML, and pushes to drafts. I open the drafts, do a quick review, and hit publish. Done in 10 minutes.

One major pitfall I hit: Python sending Chinese text to the WeChat API produces garbled output. requests.post(url, json=data) converts Chinese to \uXXXX by default, and WeChat doesn't auto-decode it. You have to manually use json.dumps(data, ensure_ascii=False).encode("utf-8"). This cost me half a day.

Twitter Operations

An intel Agent periodically scrapes X, filters content with viral potential, and pushes it to a material library. A tweet Agent picks from the materials and posts directly for me.

I don't reply to Twitter users myself anymore -- everything goes through my OpenClaw instances.

Code Review

This is what I think Agents do best for product development right now. Once configured, you can operate OpenClaw directly from your phone -- have it build websites, deploy, push to GitHub, modify code.

The full workflow looks like this: the Agent pulls project code, runs API tests, uses a browser to test page interactions, and logs issues to a Notion document. Then Claude Code makes targeted fixes based on the document, and the Agent re-checks after each fix. After a few rounds, it catches quite a few bugs that humans easily miss.

Intelligence Gathering

Multiple cron jobs monitor different data sources: OpenClaw-related content, AI going-global trends, promising new projects. Collected content automatically flows into Notion with tags and categories.

I used to spend an hour a day scrolling through feeds. Now I just open Notion and read the curated results.

Zsxq (Knowledge Planet)

Documentation Is Your Brain

There's a saying I deeply agree with: everything you do should ultimately be captured in documentation.

Whether it's gathered intelligence, pitfalls encountered, lessons learned, written drafts, or discovered bugs -- everything gets documented.

I use Notion + Obsidian for dual backup. Everything the Agents produce is automatically written to Notion (connected to 4 database APIs: work logs, content drafts, user management, knowledge base), and simultaneously synced to Obsidian as local backup.

Why dual backup? OpenClaw has a built-in memory system (MEMORY.md + memory/ directory), but network interruptions or context compression can still cause information loss. Maintain a solid documentation system, and even if an Agent loses its memory, just feed it the docs and it recovers instantly.

It's simple to set up -- just tell your OpenClaw instance "sync all content to Notion." It'll ask for a Notion API key, you give it a page's API key, and from then on everything gets synced. I basically have it write me a daily report every day, summarizing what was done and what's pending.

Obsidian's bidirectional linking is especially useful here. Every piece of information the Agent writes can connect to other notes through wikilinks, naturally forming a knowledge network over time. If you sync content to Obsidian, you can also use Claude Code locally to manage those Markdown files and Skills -- extremely convenient.

Lessons Learned

Model Configuration

At first I wasn't familiar with model configuration and kept running into issues: conversations breaking, Agents losing memory, responding but not executing actions. It took several days of tinkering to figure out how to properly configure baseUrl, apiKey, and api fields.

Now I have multiple LLMs configured as backups -- if one has issues, I immediately switch to another using the /model command. Deep thinking uses Claude Opus, coding can use Codex (free), and everyday tasks use cheaper models. I also set up failover with a watchdog that checks every 5 minutes and automatically switches to backup when the primary model goes down.

Deployment Environment

The Windows 64GB RAM main machine has been running stable -- never crashed once. I also bought a Mac Mini M4 that the team uses.

Cost

A lot of people think running AI is expensive. Let me do the math:

Hardware: one Windows main machine (already owned, 64GB RAM), plus a Mac Mini M4 for the team at 3,186 yuan. Monthly costs: electricity a few yuan, Claude Max about 1,450 yuan ($200/month, primary model), Codex free, Brave Search 35 yuan (for intelligence gathering). Average monthly total: about 1,490 yuan.

For comparison: hiring an intern costs 3,000 yuan per month, working 8 hours a day, taking days off. My AI team costs 1,490 yuan per month, online 24/7, no days off. And when models get upgraded, capabilities automatically improve.

Get Your First AI Employee in 10 Minutes

No coding skills required. No server needed.

npm install -g openclaw && openclaw onboard

Register at aigocode.com to get an API Key, create a Telegram Bot via @BotFather and fill in the Token, and you're basically up and running. Once it's running, you'll find that one isn't enough -- you'll want a second, then a third. Eventually you end up like me, managing a whole AI team.