23:56Tech With Tim · Tutorial
I Gave Codex a 24/7 Server. Now It Codes While I Sleep.
A 24-minute tutorial on running Codex CLI headlessly on a VPS so your AI coding agent keeps working while you sleep, travel, or are offline.
June 19thThe argument in one line.
MindsHub Cowork copies the Claude Cowork workflow of skills, memory, connectors, and artifacts, but decouples it from any single model provider, so a viewer can route tasks to Claude, GPT, Gemini, DeepSeek, or a fully local model without re-platforming.
Read if. Skip if.
READ IF YOU ARE…
- You already use an AI coworker tool like Claude Cowork daily and want to know if an open-source alternative can match it.
- You want the option to swap between Claude, GPT, Gemini, DeepSeek, or a local model without switching tools entirely.
- You're comfortable installing a desktop app and working through settings menus to wire up providers and connectors.
SKIP IF…
- You just want one AI chatbot and have no interest in comparing model providers or self-hosting anything.
- You're looking for a deep technical teardown of the open-source codebase rather than a product walkthrough.
The full version, fast.
MindsHub Cowork is an open-source, self-hostable AI agent tool built to feel like Claude Cowork, complete with projects, artifacts, connected apps, memory, and a skill library. The difference is a model router: instead of one company's models, you assign separate planning, routing, and coding models from Claude, GPT, Gemini, DeepSeek, or any OpenAI-compatible endpoint, including a fully local model run on your own machine through something like LM Studio. The video walks setup, a sample research-and-dashboard task, swapping in a local model, building a custom skill, connecting Asana through a credential vault, and two realistic use cases: turning a spreadsheet into a working dashboard app, and turning a brand guide into an on-brand presentation. The takeaway is that the workflow itself, not the underlying model, is what you're actually adopting.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:50
01 · Overview
Cold open on a finished dashboard built unattended by the agent, then the sponsor reveal and the vendor lock-in problem being pitched.
00:50 – 02:22
02 · What is MindsHub Cowork
Open source, self-hostable, real model router across Claude, GPT, Gemini, DeepSeek, or local models; free to use.
02:22 – 03:18
03 · Setup & Install
Three ways to run it: desktop app, self-host from GitHub, or the hosted web UI; free account gets starter credits.
03:18 – 05:00
04 · UI Walkthrough
Familiar coworker-style layout: projects, artifacts, connected apps, channels, memory, skill library; agent harness picked in Settings.
05:00 – 06:08
05 · Sample Task
One prompt produces a full interactive LLM price-vs-performance dashboard as a shareable artifact.
06:08 – 08:12
06 · Changing Models (Local & Cloud)
Adding LM Studio as an OpenAI-compatible provider and assigning a local model to the planning/coding role; swapping cloud models like GLM, Grok, DeepSeek.
08:12 – 09:28
07 · Skills
Building a custom YouTube-title-generator skill conversationally, saved to the skill library and invoked with a slash command.
09:28 – 10:50
08 · Connectors
Connecting Asana via a personal access token; credentials held in a separate credential vault not exposed to the agent.
10:50 – 11:49
09 · Memory
Self-learning lessons and rules extracted automatically from sessions, shared across the Anton and Hermes harnesses.
11:49 – 13:06
10 · Realistic Use Case #1
A mock sponsorship spreadsheet turned into a working dashboard app; a broken first pass fixed by pasting a screenshot of the bug back into chat.
13:06 – 14:33
11 · Realistic Use Case #2
A brand guide passed as a file reference produces an on-brand four-slide presentation on prompt injection.
Lines worth screenshotting.
- MindsHub Cowork mirrors the Claude Cowork interface (projects, artifacts, connectors, memory, skills) but routes tasks through any model provider instead of one company's.
- The tool splits model selection into three roles, planning, routing, and coding, so each step can run a different provider based on cost or quality needs.
- Any OpenAI-compatible endpoint, including a fully local model in LM Studio, can be assigned as the coding or planning model with zero cloud calls.
- Custom skills are built conversationally: describe the task, the trigger phrase, and the constraints, and the agent drafts and saves a reusable skill.
- A skill can research your own historical output (past YouTube titles, in this case) to build a template it reuses on every future run.
- Connector credentials are stored in a separate credential vault that the agent and underlying LLM cannot read directly; the connection authenticates without exposing the secret.
- Memory in MindsHub is self-learning: the agent extracts and stores lessons and rules from sessions automatically rather than requiring manual note-taking.
- Memory persists across the two agent harnesses (Anton and Hermes), so switching harnesses for a different task doesn't reset what the agent has learned.
- Anton is built to hand back a finished, referenceable artifact; Hermes is positioned for scheduled or automated runs rather than one-off deliverables.
- A raw spreadsheet of sponsorship data was turned into a working database-backed dashboard, not just a chart, and the first broken version was fixed by pasting a screenshot of the bug back into chat.
- Passing a brand guide as a file reference let the agent generate an on-brand presentation and hit an exact structural constraint (four slides) without extra style instructions.
An open coworker tool that swaps any model in, not just Claude's
WHAT TO LEARN
The value isn't cloning Claude Cowork's UI, it's decoupling the agent workflow (skills, memory, connectors) from any single model provider so you can swap in whatever is cheapest or best at any moment.
02What is MindsHub Cowork
- MindsHub Cowork is an open-source, self-hostable AI agent tool with a real model router, letting you run Claude, GPT, Gemini, DeepSeek, or your own local models in the same interface.
- It ships free and open source, so as soon as a new model launches you can switch to it immediately instead of waiting on one vendor's roadmap.
03Setup & Install
- You can run it three ways: download the desktop app, self-host it from the GitHub repo, or use the hosted web UI with no install.
- Signing in with a free account gets you starter credits for cloud models, or you can skip the account and bring only your own provider keys.
04UI Walkthrough
- The interface intentionally mirrors familiar coworker tools: projects, artifacts, connected apps, channels, memory, and a skill library.
- Under the hood it runs on a swappable agent harness, Anton or Hermes, chosen from Settings > Agent.
05Sample Task
- A single prompt produced a full interactive web dashboard as a shareable artifact, with password-protected or fully public link sharing.
- Anton is built to hand you a finished, referenceable artifact; Hermes is the harness meant for scheduled or automated runs instead of one-off deliverables.
06Changing Models (Local & Cloud)
- Any OpenAI-compatible endpoint, like LM Studio or Ollama, can be added as a provider so a fully local model can run the planning, routing, or coding role with zero cloud calls.
- Planning, routing, and coding are three separate model slots, so you can pair a cheap fast model for coding with a stronger one for planning instead of paying premium rates everywhere.
- Swapping providers is a settings change, not a re-platform: the same app can go from a paid frontier model to a free local model in under a minute.
07Skills
- Skills are built conversationally: describe what the skill should do, when it triggers, and what outputs it needs, and the agent drafts and saves it to a reusable library.
- A custom skill can research your own back catalog to build a template it reuses on every future invocation, not just a one-off output.
08Connectors
- Connectors cover the standard SaaS stack (Gmail, Google Drive, Slack, Salesforce, Asana, etc.), authorized once via a personal access token.
- Credentials are stored in a separate credential vault the agent and underlying LLM can't read directly; the connection is used on the app's behalf without exposing the secret in context.
09Memory
- Memory is self-learning: the agent extracts and stores lessons and rules from your sessions automatically rather than requiring manual journaling.
- Memory persists across harnesses, so switching from Anton to Hermes for a different task doesn't reset what the agent has already learned about you.
10Realistic Use Case #1
- Handed a raw spreadsheet of mock sponsorship data, the agent built a working database-backed dashboard rather than just a static chart of the sheet.
- The first version had rendering bugs; pasting a screenshot of the broken UI back into chat was enough for the agent to diagnose and fix it on the next pass, normal iteration, not a one-shot result.
11Realistic Use Case #2
- Passing a brand guide as a file reference let the agent generate an on-brand interactive presentation without additional style instructions, while still hitting an exact structural constraint.
Terms worth knowing.
- Model Router
- A layer that lets one agent tool call multiple different AI model providers interchangeably, instead of hard-wiring the tool to a single company's models.
- Agent Harness
- The underlying execution engine that runs an AI agent's tasks. MindsHub ships two, Anton and Hermes, switchable from the same account and memory.
- Credential Vault
- A storage layer that holds API keys and access tokens for connected apps so a request can be authenticated without exposing the credential to the AI model or agent itself.
- OpenAI-compatible provider
- Any model server, local or cloud, that exposes the same request format as OpenAI's API, letting tools built for that API (like LM Studio or Ollama) plug into any agent app expecting it.
Things they pointed at.
06:22toolLM Studio
08:46toolAsana
Lines you could clip.
00:06
“You're locked to one company's pricing, one company's ideology, and one company's models.”
01:46
“Rather than running everything through something like Claude Cowork, where you're completely locked into Anthropic's environment, you can use an open source harness... and switch out the model whenever you want.”
10:19
“All of the credentials that you put here go into something called a credentials vault... it's actually not viewable by the agent or the LLM itself.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphor
This dashboard was built by an AI agent called Anton. I briefed it once, walked away, and this is what was waiting for me. Now, I use Cloud Code and Cloud Cowork every single day, but it has one fundamental problem, and that's that you're locked to one company's pricing, one company's ideology, and one company's models.
Now, the tool I use to build this dashboard is called MindsHub Cowork. Now, it's fully open source. You can run it hosted or you can self -host it on your own machine.
It has a real model router, so you can run Claw, GPT, Gemini, DeepSeek, or even your own local models. Now, they sponsored today's video, and I'm going to show you exactly how to set it up and use it start to finish. The key differentiator here is that it's free.
You don't need to pay for it. Again, it's completely open source, so you're in control. And as soon as a new model comes out, you can immediately start using it, and you're not locked into one company's pricing.
Anyways. Let's dive in. So here's a quick look at Mindsub co -work.
You'll notice it's a very familiar user interface. Looks like many of the other co -work applications. You have projects, artifacts, connected apps, channels, memory settings, skill libraries.
You get the idea and then you can click into a task and obviously interact with the model and see what it's produced and the progress that it has. Now, again, the main differentiator here is that this is fully open source. That means that there's multiple ways to run this application.
You can download the desktop app, you can self -host it yourself, or you can even run it in the cloud directly from MindsHub, which I'm going to talk about in a second. Now, in my case, I'm just using the desktop application because it's the easiest to set up. But the reason you would even use a tool like this is because you have all of the options.
Like I said, it's open source. You can play with the code on your own and you can choose any model you want at any point in the future. And that's super important now because we've seen Kimi's come out.
We have GLM. We have all of these great models, which are pretty much as good as something like fable, but are so much cheaper.
So rather than running everything through something like Claude cowork, where you're completely locked into anthropics environment, you can use an open source harness, like we're using right here and switch out the model whenever you want. And of course you can still use cloud models, but you can also run local models and you can choose any provider at any time with no vendor lock -in, which is the reason I'm even showing this to you.
So anyways, what I'm going to do is quickly show you how to download it and set it up. Then what I'm going to do is walk through the process of adding skills, connecting data, and how you use this just like an agent coworker, except with all of the benefits that I discussed. Now, the easiest way to get started with this tool is to download the desktop application.
So from this page here, I'll leave a link to it in the description. You can just press the button that says download for your operating system. So let's start with that.
Now we'll run through the installer in a second, but if you do want to self -host this, you can go right to their GitHub. All of the code is open source. You can download it, run it, do whatever you want with it directly from here.
Also, you can open it directly in the web. They have a web UI. If you just want to do it from here and you don't want to download anything.
Okay. So let's run through the installer and let me show you how we get this set up. Okay.
So the installer's finished. And from here, you can either continue without an account. If you just want to do everything yourself and create your own models and set everything up.
However, if you just want to use this super easily, then you can connect to your mind's account. So you can create a free one or sign into it. And then from there, you'll get a bunch of free credits to use some of the cloud models, or you can of course bring in your own providers, which we'll show you in a second.
So from here, if you don't have one, create a free account. In my case, I'm going to sign into my account. Okay.
So I've signed into my account here and now I can start using this just like any cowork tool. Now behind the scenes, what this tool uses is something called Anton. Now Anton is an agent harness, somewhat similar to something like Hermes.
Okay. If you've seen like Hermes agent before, and actually within mindset cowork, you can change between the harness that you want to use. So if I scroll here and I go into the settings and I go agent, you can see that I can select the harness of either Anton or Hermes.
So I can go between them. and then save the settings if I want. Now Anton is the best default option because it's the best for actually providing usable output.
Whereas Hermes is more for like running things automatically on a schedule. Now here I'm connected to Mindsub already because I've signed in with my account. That means that I can use any of the cloud models.
You can see they have a bunch of ones here like Mindsub Air. This is a free model they provide that gives you a ton of additional usage, or you can pick any of the other models like GLM 5 .2 or Grok 4 .5, right? Or Kimi K3 or whatever you want, and you can choose what you want for planning, for routing, and then for coding.
Now, if you don't want to use this, you can connect to your own provider. So Anthropic, OpenAI, whatever, or use any OpenAI compatible API. So something like LM Studio or Ollama.
And if you want to see your usage here with Mindsub, again, you'll get a bunch of free credits. And then if you want to keep using their model router, you can, of course, like buy a subscription, buy additional credits, et cetera. The way you would see that is from the website.
You can go to your console. From the console, you can go to cowork. You can top a for using the cloud models, or you can get a ton of additional credits again for their minds hub air model, as it describes here.
Okay. So let's go back and actually for the coding model, I'm just going to switch this to GPT 5 .6 SOL or soul for now. And let's save that.
Okay. And then let's just give it a task. So I'm just going to ask it something simple.
Can you do research on the top LLMs right now and create a dashboard that goes over the performance versus the price? I want you to cover at least 10 different LLMs, go through the model families and give me a ton of information and an interactive dashboard that I can view. Okay.
So let's do that. And then go ahead and press on enter. And I just want to show you what it generates and how it works.
Then we'll go into all of the other settings. Okay. So just wrapped up.
Now we can see, we've got all of the info here and we can actually go to the artifact that it generated. which is a full web dashboard, and we can view everything inside of here. And then if I want, what I can do is actually just share this.
I can add a password, I can select users, or I can make it fully public. And then anyone is able to view this. Now, this is kind of the difference between Anton and the Hermes, um, harness that you have here inside of mind sub with Anton, it's meant to produce these artifacts that you can then reference later and actually give you like fully finished work.
Whereas with Hermes, it's typically better for automations, running things like on a schedule. get the idea and then kind of that skill curation so anyways if we weren't happy with this of course we could do a follow -up prompt and we could make it give us something better but for now i want to show you a few other features for example i want to show you how we can change to use something like a local model so for example if we go into these settings and we go to the models here we can add another provider so if we go to provider we can go open ai compatible what i can do is just call this like LM studio, for example.
And then if I pop up an LM studio, let me just load up a model here. So let's load like Java for, for example, let me just pump up the context window here in a second. And then what I can do is find the API end point.
So it's just this one right here. And what I'm going to do is just reload this quickly. So it gets the updated context from here.
I'm just going to put nothing for the API key. And I'm going to put this slash V one, and then we're just going to wait a second. So once LM studio loads this up and I can save.
this. And then what I'll be able to do is actually adjust the model here in just one second. Okay.
So we can see this provider is added. So for example, if I want my planning model to be LM studio, I changed the provider. And then for the model ID, I just have to copy it here, which is Google Gemma four.
So let's paste that right there. And then I can save the settings and now I've adjusted this. So for my planning model, if I go and I make a new task, I can say, Hey, what model are you using right now?
And if we hit enter, it should use the planning model and we should get the LM studio response. And there we go. We get it.
Gemma three, 4 billion, just using a small one. And you can see how fast it is again, because I'm running that local model on my own machine. Now you can switch to any provider that you want, but I'm just trying to indicate that, you know, it's very good, easy to switch and to use local ones directly in this harness.
Okay. So for now, let's go back and change the model because this one's not going to be quite as good as what I want for these examples. Let's see, what do I want to actually use here?
Um, let's try like deep seek V4, for example, reasoning level, we can go with high. We'll change the routing one here to use mind sub air, just because we don't need super high intelligence for this one. And for the coding model, I want to go with something a little bit faster.
So let me see if I can find grok here, which is pretty fast. So yeah, let's use grok 4 .5. Okay.
So we're going to just save those for right now. And then you'll notice that there's a bunch of other stuff we can do here.
Like there's a backend, the account channels, all this stuff. I'll show you how we set that up in a minute. Okay.
So now that we've got that, let's have a look at adding skills, for example, because similar to kind of Hermes, there's a skill library here and you can manually create a new skill by asking cowork. You can write the instructions or you can upload one. So let's just do a very simple skill.
Let's create a skill together. First, ask me what the skills should do. Let's hit enter.
I just want to show you how that works. So just to do a super quick skill here, let's do something like this. I want to create a skill that can generate a YouTube title for me.
The idea should be to generate something that peaks curiosity and interest. Overall, we want to keep it under 100 characters in length. We want it to be to the point and we want it to be kind of tutorial based and full course based.
Go look at my YouTube channel tech with Tim. Find some of the titles that have performed well and then build this skill around that so that we always get titles that kind of fit that narrative. Not sure if that makes a hundred percent sense, but let's tell it to go do that and see the skill that we get.
All right. So we can see it actually went, it looked up a bunch of my YouTube video titles, saw how they were performing, gave me a bunch of stuff here, output. And then I can just save this skill directly.
And if I go to the skills library, now we now have the YouTube title generator that was written. Now, if we want to invoke the skill, of course, as usual, we can just go slash, and then we can just do the skills. So YouTube title generator, and then just ask it what to generate, and it will create it for us.
Now skills are great, but we also can connect to different applications. So if we go to connect here, there is a ton of connectors that are built in.
So for example, we've got Gmail, Google Drive, Salesforce, Slack, right? You name it, pretty much all of the integrations are here. Now, one that I'll connect to for this video is Asana because this keeps track of like my video pipeline and stuff so I can show you a demo.
So let's go here. I'll grab my personal access token and then show you the connection. Okay, so I've got the token.
I'm now just authorizing it here. So I've given it to Anton. And in a second, what it should do is start connecting this and then allowing me to talk with Asana.
Now it's worth noting that all of the credentials that you put here go into something called a credentials vault. This means that it's actually not viewable by the agent or the LLM itself, but it will be used when it actually needs to access the service. So it's automatically handled and that's kind of part of the software.
You can look at it obviously in the open source code if you want, but it's called the . So it everything secure.
All right. So it looks like we're connected here. We've got the connector.
And if we want, we can go to a new task and say, give me a quick summary of what I have coming up for videos in a sauna. And this is because it tracks all of my video status, what needs to be edited, kind of like the pipeline there. So let's see what it shows us.
Okay. And you can see that it's pulled all of this information to videos that are being edited. And then a whole sequence of shorts that I'm doing for a new school community that I have, by the way, I'll leave a link to the description if you want to build AI agents yourself, but you can see that it pulls everything up.
Nice. So I think the last thing to mention here is the memories. So like pretty much all of these agents that can pull from the session history, but also it can automatically keep track of rules.
You can add stuff in the profile, for example, you can give it lessons, you get the idea. So if you want it to always remember something, just tell it, it can automatically update this, or you can edit the file yourself, add the markdown and it will remember. Now, just to clarify here, the memory that you see is self -learning.
That means that it's going to automatically remember what you tell it. And these files that you see here aren't really designed to be edited manually. It's more so that you can view them and you can see what co -work has extracted and remembered.
Now, another thing to note is that the memory is shared across harnesses. So that means that you can easily switch between Hermes and Anton and you keep the exact same memory, making it a lot easier to switch between harnesses. Okay.
So that's kind of a general walkthrough. Of course we have the scheduled tasks too, but I'm not going to show that one right now. What I want to do is show you a few legitimate use cases of what these types of tools can produce and why you would actually use them.
Okay. So for the first example, I have this kind of Google sheet with a bunch of information, just like fake sponsorship data, right? So what I'm going to do is I'm actually going to have Anton convert this into a fully functioning dashboard and web user interface.
And that's what this prompt does. It turns it into a legitimate database application that I. could actually deploy, which is going to be a lot more visual than something like the sheet.
So let's run this and let's see the output we get. Okay. So it took a few minutes here, but it looks like we've got the result.
Let's check it out. And okay. You know, that's not exactly perfect.
It looks like there's a few issues. So what I'm going to do is just copy some of this and I'll just tell it to fix this up. And look guys, sometimes that happens when you're using AI agents, it's not always perfect.
So let's give it another prompt to fix that up. Hey, there are some pretty big issues. The UI is not really rendering properly.
It's all kinds of random text on it. Can you have a at this again and redo it okay and i'm just going to paste this in this is an example of what i'm seeing on the actual screen again the ui is pretty messed up okay so just wrapping up now but if i go to the artifact we can see now it's working there we go that's better that's what i was looking for and we get the full dashboard with all of the deals we have any payments that came in deliverables sponsors so we effectively just took the google sheet and converted into a real legitimate dashboard that's actually usable nice let's go to another use case So something I've been doing a lot recently is actually making interactive presentations to go along with some of the videos that I have.
So for example, I have this brand guide for my new school community called AIH of Builders. It has all of the information about it. So what I'm going to do is just pass this in as a file reference here to MindsHub or Anton.
And I'm going to ask you just to make a quick presentation about something, right? And then use this brand guide. So let me come up with a prompt and I'll be right back.
Okay. So I just asked it to make something about prompt injection. Let's see what it can do and the result we get.
Okay. So just wrapped up here. If I go to the artifact, we can see this is what we've got.
Prompt injection follows the guidelines. Exactly. We have, let me make it actually full screen here so that we can see it a little bit bigger.
Okay. There we go. You can see, we have kind of little box down here.
We have the code snippets. We have real world impact defense layers, and it's exactly four slides, just like I asked for. Now again, it's not super complex with the point is it follows the instructions and I can add.
these documents and create something really meaningful. Now look guys, there's so many other things that you can do with tools like this. I'm going to leave it to your imagination.
The main reason why I like a tool like this really is because I have full flexibility with all of the different models and the providers, and I can keep all of my data and information local and really have full control. So anyways, that's going to wrap it up. If you guys enjoyed the video, make sure they like subscribe to the channel and I will see you in the next one.
The Hook
The bait, then the rug-pull.
The video opens on a finished dashboard, not a pitch: an AI agent named Anton built it unattended, and only after landing that payoff does the video name the sponsor and the vendor lock-in problem it claims to solve.
Named ideas worth stealing.
02:35concept
Three-slot model routing (Planning / Routing / Coding)
- Planning model
- Routing/summarization model
- Coding model
MindsHub splits the agent into three separately-assignable model roles instead of one fixed model, so each role can run a different provider based on cost or quality tradeoffs.
Steal forany multi-step agent tool where planning and code generation have different quality/cost requirements
How they asked for the click.
VERBAL ASK
00:27product
“Now, they sponsored today's video, and I'm going to show you exactly how to set it up and use it start to finish... it's completely open source, so you're in control.”
Sponsor disclosure delivered inside the opening pitch (first 50s) rather than as a separate ad-read break, framed as the reason he's showing the tool rather than a detour from it.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
