Episode 111
THE AI THAT ACTUALLY DOES THINGS: OPENCLAW, DISCORD, AND GEMMA 4
What happens when you give an AI the keys to your terminal? In this episode of Venture Step, Dalton Anderson dives into the agentic revolution with OpenClaw—the open-source powerhouse that's so disruptive it…
Episode content
Episode Story
OpenClaw: Gateway, Channels, Models, Agents, and Tools
OpenClaw is a self-hosted personal-agent gateway that connects messaging channels to AI models, sessions, agents, tools, nodes, and automation.
What a Messy OpenClaw Setup Reveals
A real OpenClaw, Discord, Ollama, and Gemma 4 setup shows why identity, secrets, model capacity, supervision, and recovery matter more than installation.
Gemma 4: Models, Capabilities, Context, and Local Use
Gemma 4 is Google's open-weight multimodal model family, with several architectures and sizes for edge devices, laptops, consumer GPUs, and workstations.
Discord: Servers, Channels, Bots, Permissions, and Intents
Discord is a communications platform with servers, channels, threads, apps, bots, OAuth2 permissions, and gateway intents for real-time integrations.
Research & Analysis
Field Notes
What Is OpenClaw? Gateway, Models, Channels, and Tools
OpenClaw is a self-hosted gateway that connects messaging channels to AI models, sessions, agents, and tools. Learn how its control flow and permissions work.
OpenClaw Security Checklist Before You Connect Anything
Secure an OpenClaw experiment with isolation, least privilege, pairing, protected secrets, restricted tools, logs, backups, audits, and a tested shutdown.
How to Connect OpenClaw to Discord Without Guessing
Connect OpenClaw to a private Discord bot with least-privilege permissions, protected tokens, pairing, allowlists, text-only tests, and ordered troubleshooting.
Local Model vs API Model for a Personal Agent
Compare local and hosted models for personal agents by task success, privacy, latency, reliability, tool use, hardware, maintenance, and total cost.
Full episode
TranscriptSearch or read the full conversation.
E111B The AI That Actually Does Things: OpenClaw, Discord, and Gemma 4
Transcript
Dalton Anderson (00:05.784) Hey, I'm back. So actually we spent time downloading something we didn't need to download. There was an error in the documentation I created and I just copied paste and it was downloading the 26 parameter, 26 billion parameter model, which is not what you want to download on this type of machine. I'm running an M1 Mac with 16 gigs of unified memory. So basically no shot. I could run that.
I can't even run a 12 billion parameter model. So the four billion parameter model is what I'll be running, which it's supposed to be really good for the.
Dalton Anderson (00:48.528) the amount of space it takes up on your RAM and how small the model is. So just keep that in mind. I used it last night. It seemed okay, but I was trying to do complicated stuff apparently. But at first simple task, it seems fine. So on the right side here, I've got two terminal windows open here. On the right side of the window, you'll see that I've got this command to run olama run gemma.
and it's gonna test if it's online. Let's see, this doesn't exist. All right.
Let's see here, Oglala operational. I swear I just ran it.
Dalton Anderson (01:38.012) say test.
Dalton Anderson (01:43.708) wanted to run it different screen, but then I had to change the command a little bit.
Dalton Anderson (02:03.886) online and works.
Okay, so it's gonna test it, it works. I was looking at it earlier and I was asking questions about the model, because it downloaded the 26-point parameter model. I was like, hey, yo, what's going on? I tried to cancel it and I was like, that's way too big. I can't run that on my machine. And then it said that it was up and running, so I did a test, said hello, and it said, hello, how are you, how can I help today?
So then it just said, test successful, online, working perfectly well. I am online and functioning optimally, ready to assist with any requests. Your setup is now fully verified and ready to use. Is there anything else you'd like us to do? OK, so the next thing that we want to do is we want to onboard.
We want to onboard open claw. you just open claw and then onboard is all you have to put. And then it's going to ask you a whole bunch of questions when you onboard it, which is important because this is when you gather your keys. So it's going to ask you, okay, how are you going to interact with it? What are you going to do? Et cetera, et cetera. And from that, you'll need to give it keys and information.
Uh-uh.
Dalton Anderson (03:37.404) I hope it asked me because it's important part of the video, but it may have taken it from my old setup.
Dalton Anderson (04:00.474) agent's main session.
Dalton Anderson (04:08.059) Well, it didn't ask me anything. So how can we reset up?
Right, one.
Dalton Anderson (04:24.642) change.
Dalton Anderson (04:49.083) I spelled ground wrong. Grun. Let's see. But it would ask you, hey, are you okay with this? Who can we trust here? It asked similar questions earlier, but.
Dalton Anderson (05:06.139) Okay, so it moves everything to the trash. And that's gonna restart the session. And then it's gonna ask me questions that I'll have to answer.
Dalton Anderson (05:34.043) Okay, so we will be configuring and then we'll be setting up the discord and then we'll just test it and then the show is over.
Dalton Anderson (06:08.283) Exploring, exploring, exploring.
Dalton Anderson (06:27.333) I always think it's really neat when you get to...
Too long, save.
Dalton Anderson (06:54.235) What a pain. It seems like it's harder to set. That's a weird thing in coding. It's always harder to set it up again once you've already set it up because there's always files that stay on your machine even if you wipe your machine, not necessarily wipe it, but wipe your profile and delete.
Dalton Anderson (07:27.261) there I have to do my bad.
Dalton Anderson (07:34.073) you just running here
Dalton Anderson (07:44.623) It should ask me all the questions that I'm talking about. All right, here we go. Perfect. Let's make this big screen. OK, so this is the setup. This actually looks worse.
Just
like this. It should be fine. OK, so I understand this is a personal by default. And I say yes. And then I'll say quick start. And then I want to update values. OK, so which model provider are we using? So this is what I was saying. There's a lot of choices. We are not using these other ones. We're going to use Olamma. But these all have
Different choices. mean, there's a lot. Figure out what's good for you. For this instance of the show, I'm going to use Olamma. But probably going forward, I'll be using Gemini command line. there you go. Authentication. OK, so Olamma. So it starts up.
Dalton Anderson (08:55.099) And then we will do local or you could do cloud and local. So I'll just be using local.
Dalton Anderson (09:10.657) Okay, so it's going to ask me what model I'm going to use. And this is the model that we downloaded earlier, GIMMA 4. So let's take that.
Dalton Anderson (09:23.343) This is not found, that's lie. So this is how you can going to communicate with it. So there's many options. You could do Google Chat, the bot framework from Microsoft Teams, Blue Level is what I talked about earlier. Then there should be right here, Discord. And so click Discord and you're have to give this your bot token. So enter the bot token.
So we're go back to our other webpage.
and then we're going to get our bot token.
which I put Discord bot token right here. Then we'll give it.
I didn't copy and paste that correctly probably, so let's just redo that.
Dalton Anderson (10:19.812) Okay.
Dalton Anderson (10:24.988) Guess I didn't copy and paste it. Correct again, twice in a row. Crazy. Copy.
Dalton Anderson (10:37.852) based.
Configure bot channel access and I'll just say allow list. I wanted to say allow off.
Dalton Anderson (10:53.724) All right, let's just redo it. I don't know how to go back.
start.
Dalton Anderson (11:08.412) Olamma, enter local. Then it loads for a second. Then I'll select the model.
So then we'll say Gemma.
Then after that, we're going to use Discord. It's down the bottom.
Dalton Anderson (11:30.17) Enter Discord token.
Dalton Anderson (11:37.166) my screen just went black. my gosh. Weird one. Okay, I'm back. Configure channels. And my screen just went black again for some reason. What's going on with my Mac? I say allow all. Why am I saying allow all? Because no one else is in there besides the bot, so I don't care. And then...
Dalton Anderson (12:03.382) I planned on using...
Dalton Anderson (12:10.405) this model.
Dalton Anderson (12:14.818) API key.
I'm going to stop sharing for a second. Stop sharing.
and then I'll give it the key because I don't want to give it the key and share it. I don't want to have to make a new key.
Dalton Anderson (12:33.605) So let's do that.
Dalton Anderson (12:42.971) Figure skills now. All right, let's go back to sharing.
Dalton Anderson (12:50.595) It's fine, I don't need web search. Just skipped over it. Share window.
screen.
Dalton Anderson (13:05.851) my gosh, so freaky. so configure skills now, yeah. And then you can add in all these skills if you want. I'm not gonna do that for this. We don't need to do that, so skip. You can edit that later.
Dalton Anderson (13:40.859) You could set up these things.
Dalton Anderson (13:46.53) and then restart.
Okay, so now.
Dalton Anderson (13:55.408) You can keep going through this process, but now we need to run it.
Dalton Anderson (14:09.433) we
Dalton Anderson (14:20.571) Get that going.
Dalton Anderson (14:40.827) Okay, testing.
Dalton Anderson (15:04.635) clock, gateway start.
Dalton Anderson (15:09.177) Hmm. Let's see here.
Dalton Anderson (15:32.355) All right. So then the other thing is you can look at it.
And then we should be able to interact with it using this. Let's say testing right here.
Dalton Anderson (15:57.532) And it's not there. So it's not online. It's a good way to tell whether it's working or not working. So let's see here. I'm just gonna copy. Sometimes when things are desperate, you just ask and oh, here we go. It's online. Cool. Hey, I'm awake. Wake up my friend. Hey, I'm awake, ready to help.
Dalton Anderson (16:31.651) I'll be with your testing today. All right, cool. I was like, this is taking forever. OK, so then.
Dalton Anderson (16:43.279) Discord, we're get the Discord going. And then we're finished, okay?
Dalton Anderson (16:54.222) Are you?
Dalton Anderson (17:10.615) And I don't mind chatting to AI while we're on here because I'm just showing how not complicated this is. It's just, you gotta ask questions and be wrong, try again, and just keep attacking. Just keep attacking. And I think we already did this on my other machine, but I hadn't run this yet, so, but this is important. This is the sandbox part.
And when the AI dies out, it just will restart.
Dalton Anderson (17:53.711) Yeah, I've got the command token.
See what AI can book up. AI can, AI AI? AI AI'd. Let's see.
Dalton Anderson (18:24.443) Okay, then the next thing that we need to do is we need to...
back over here and take this.
Dalton Anderson (18:52.025) Message content, yeah.
Dalton Anderson (18:57.859) Let's see here. We need to go back and ensure this is done correctly. It is correct. I got it.
Dalton Anderson (19:36.091) It's not showing up. Weird. It's always flawless when you're doing it by yourself.
So funny. So funny. Like last night, the only part of my setup that took a while was restarting my machine and then getting this going was easy. It took me 30 minutes.
Dalton Anderson (20:08.667) but this is not not vibin'.
Dalton Anderson (20:20.059) But it's probably because I skipped the PM2 setup on accident.
and then that manages the agent and makes sure that if it crashes it restarts and it does a whole bunch of stuff.
Dalton Anderson (20:45.483) We'll get it right. Don't you worry about it. We'll get it all right. We'll get it all right. We've got all the stuff that we need in here done.
Dalton Anderson (21:10.693) Got this set up, so.
Dalton Anderson (21:17.721) We should be fine. We should be fine. Just gotta get this thing to work.
Dalton Anderson (21:26.629) And then after it's working, then I'm going to bed.
Dalton Anderson (21:32.571) I'm tired, tired of AI. I set this up last night and I thought it wasn't working because it took so long for the AI to reply back to me. Like it took like four minutes. It was typing for four minutes and I just said, hey testing. I asked it, why did it take you so long to reply? And it said something along the lines of, that's how long it taught me to.
That's how long it took me to think about what to say. I asked you how you're doing. Like are you here? Can you reply to me? It's like a yes, no. And that basically said yes, I am available. That's what it said. It didn't say anything complicated. It's so funny.
Dalton Anderson (22:24.621) So, so funny. Why this is going, let me run. I think it's Claude.
Dalton Anderson (22:44.707) Hey, I'm setting up OpenClaw on my computer and it says it's initialized and is ready to help, but then it doesn't say anything in my Discord and I am confused. So I want to run some scripts to test it on my local machine to make sure that everything is working. Do you have any suggestions?
Dalton Anderson (23:15.385) status all okay.
Dalton Anderson (23:28.889) to probably see the status of the Discord probably is the most important communication channel.
Dalton Anderson (23:41.145) While that's stinking on there, I'll have this stinking over here.
Dalton Anderson (23:48.827) Everybody's working. Everybody is working right now.
So funny though, because when I set this up, flawless, first time. First time. It just wouldn't respond to me. And then I had to add that setting about the message intent.
Dalton Anderson (24:21.493) Somehow I've already used 14,000 tokens. No, 14. Yeah, 14.
I guess this guy's running some stuff on the side, whoever.
Dalton Anderson (24:44.038) Come on,
Nodding.
Not sure what that means. think it's just looking for an answer.
Dalton Anderson (26:11.044) Okay, so then the next thing that I want to do is take this.
Run this in my terminal.
Dalton Anderson (26:27.42) Wait, wait, hold on. Is reporting enabled configured?
Dalton Anderson (26:42.202) All right, that's my bad. That's my bad. I saved the bot token earlier when we were chatting and I was like, well, then I'll just have to redo it. But now I just have to redo it anyway. So it doesn't matter. So I'll just do this. So let's say, OK, what happened was I didn't want to create this because this what I'm going to create as my actual new bot. And I was like, OK, well.
If I do that, then I have to make a new one anyways. So let me just create VentureStep because this is related to VentureStep. Oh, you'd see that. Testing Cloud connection, local script successful. And so then I created a new channel live and then created another bot live so I wouldn't have to create something again. But whatever, I'll just make a new one. OK, so we'll say hello.
and you hear me.
What is the weather like in NYC?
Cheers.
Dalton Anderson (27:58.433) wow, it dropped an emoji?
That's new.
I dropped an emoji. AI just dropped an emoji on me. Thinking face. Okay. All right. You got that swag. You got that swag. Drops an emoji on me. my goodness. It's probably going to tell me to look it up or something really mean. I'm telling you. The other AI that I created last night was so mean to me.
I was like, wow, you've got a lot of friends, don't you?
and you're so nice and this is how you talk to me.
Dalton Anderson (28:47.324) disappointed.
I wonder if I can stall for two minutes while this is thinking about replying, because it's going to take a while.
Dalton Anderson (29:00.774) do do do do do do do do do do do do do do do do do do
Dalton Anderson (29:12.966) Yeah, I don't have much, mean, it's getting late, my time on this recording session, long session, but I thought it'd be a cool episode given that this is a new technology and I'm exploring it with you, my God. I hope that it's productive, who knows if it is.
One thing that does concern me is the latency between message and reply.
Dalton Anderson (29:41.116) I feel pretty strongly that setting this up to be API based is probably the way to go if you want to be super productive. Because if you've got to wait like five minutes for it to reply every time, that's a big latency. When you think about it, if you're 10, just call it 10 messages a day, that's 50 minutes. Okay. Or so now you start sending, you start sending
20 a day. That's a long time. That's too long.
So if you can get that down to 20 seconds, 15 seconds, well then now you're cooking. It just added a new emoji. Now it's fearful.
Why is it fearful?
Dalton Anderson (30:34.095) Why is it fearful of me? AI is now fearful. It was thinking and now it's fearful. This is not good.
Dalton Anderson (30:46.811) Not sure, it just stopped typing too, it just stopped. What? Are you, hey, do you plan?
Dalton Anderson (31:11.259) I don't know. I don't know what happened to him.
Dalton Anderson (31:39.855) I don't know, it's just not replying anymore. It's here, it's running.
Let's just test it and let's run this broadcast.
Dalton Anderson (31:59.899) Ahem.
here we go.
Dalton Anderson (32:08.549) the model did not produce a response for auto timeout. Please try again or increase. Agency fuel timeout time and you can pick. Disable, set to zero, disable. got it.
Dalton Anderson (32:28.783) Alright.
Dalton Anderson (32:32.731) I think the last thing I want to show you is OpenClaw.
What is the...
Dalton Anderson (32:47.931) I'll just look it up, I guess. What is the op-
Dalton Anderson (32:53.051) local.
So that there's so much stuff going on in this that you can look at it at a dashboard.
can open this in your web browser.
All right, so then let's do that. Shimmer screen. I think it already attempted to do it. Here go.
Dalton Anderson (33:19.919) gateway token mismatch. Huh. Okay, well it's not working and I don't want to.
Dalton Anderson (33:32.845) in the skyway.
Dalton Anderson (33:37.179) I don't know if this is raindrops.
screenshotting on here is so hard because it's, it be easier just to print this out. Print. Okay, let's see.
Sure.
Dalton Anderson (34:05.538) Everything was such a battle today. It's crazy.
Dalton Anderson (34:12.923) Hey, the Claude is saying, not Claude, Claude is saying that is having issues connecting me with this dashboard. How can you help me?
Dalton Anderson (34:35.333) to do do do do. Closer.
Dalton Anderson (34:58.747) So let's do that.
Dalton Anderson (35:30.267) I'm so lazy.
you
Dalton Anderson (35:54.107) All right, let's close this, terminate.
Dalton Anderson (36:04.54) Yeah, because we never ran it in PM2, so let's do that.
Dalton Anderson (36:15.227) just want to show you the dashboard.
Dalton Anderson (36:26.331) because it's pretty cool.
Dalton Anderson (36:33.029) my opinion at least. I was like, wow, that's...
quite nice.
Dalton Anderson (36:51.003) really need to learn the Mac commands. You can tell that I use.
it still has this thing. Unauthorized gateway.
Dalton Anderson (37:14.853) All of these things that are occurring live did not happen.
Dalton Anderson (37:40.527) Okay, we're just gonna call it. It's not working. I'll edit this out, it's fine.
Dalton Anderson (37:48.955) Let's do this.
Dalton Anderson (37:58.383) Okay. Well, this was the install and overview of open claw and it was exhilarating error, error filled. never have as many errors as you do when you're doing something live. It's the live demo curse. It just comes with the game, but we took you from zero to hero ish.
Still a long way to go from being productive, but it has potential, right? It's close. So that being said, appreciate you wherever you are in this world. Good evening, good afternoon, good morning. Thanks for listening and listen in next week. Goodbye.
SourcesFollow the evidence trail.
E111B Sources
[[E111B Transcript - Dalton]] contains the first setup session and Dalton's comparison of messaging channels, model choices, and deployment options. [[E111B Transcript - Setup Demo]] continues the installation, local-model test, Discord configuration, and troubleshooting. Both raw transcript notes remain unchanged.
The transcripts control Dalton's actions, observations, and conclusions during the recording. They do not control current OpenClaw, Discord, Ollama, or Gemma behavior.
Current product sources
OpenClaw release 2026.7.1, commit 2d2ddc4, is the pinned source for version-specific package and documentation review. The package identifies itself as version 2026.7.1 under the MIT license.
The OpenClaw getting-started guide controls the current prerequisites, installation, onboarding, provider configuration, and first-run sequence. The official Discord guide controls current bot, pairing, permission, channel, routing, and troubleshooting behavior.
The OpenClaw security guide is required reading before any public setup instructions. It documents allowlists, session isolation, secret handling, network exposure, and the limits of OpenClaw as a multi-user security boundary.
The current pairing guide, session guide, secrets guide, local-model guide, and Ollama provider guide control their respective surfaces.
Discord's application quick start, OAuth2 and permissions guide, OAuth2 reference, Gateway documentation, and Developer Policy control bot identity, tokens, scopes, permissions, intents, and platform rules.
Google's Gemma documentation, Gemma 4 model overview, and Gemma 4 model card control current model variants, architectures, modalities, context, licensing, and intended deployment surfaces.
Ollama's Gemma 4 library record controls its current packaging route. It is a runtime distribution source, not the authority for Google's underlying model claims.
Google's current Gemini API pricing page was reviewed for the hosted-model cost comparison. Public copy avoids fixed price claims and requires a publication-time refresh.
Local verification record
The OpenClaw repository was cloned and pinned to release 2026.7.1 at commit 2d2ddc43d0dcf71f31283d780f9fe9ff4cc04fe4. Source documentation and package metadata were reviewed from that checkout.
A clean CLI attempt used a new temporary OpenClaw home with no user configuration or credentials. The package correctly refused to run because the editorial host had Node 24.14 and release 2026.7.1 requires Node 24.15 or later in the Node 24 line. The check changed no account, channel, token, or persistent OpenClaw configuration.
The runtime gate prevented a live dry-run configuration test in this environment. No unsupported Node bypass was used.
No Discord application, bot token, server, provider account, model download, external gateway, or production process was created for the editorial review. The Discord Tutorial is source-verified against the pinned release and current official portals and explicitly discloses the absence of an end-to-end external account test.
Evidence boundaries
The transcript records a real but unstable setup session with discarded configuration, a model-size mistake, token handling, gateway problems, Discord troubleshooting, and long local response times. It should not be converted into a copy-and-paste tutorial.
The public package must never reproduce API keys, bot tokens, local paths, account identifiers, gateway tokens, or secrets. It must not recommend allow all, disabling timeouts, public network exposure, or broad tool permissions merely because a Discord server is private.
Dalton's M1 Mac and selected Gemma configuration provide one user experience, not a general benchmark. The setup reached a limited “zero to hero-ish” state, not a verified production deployment. Current guides must distinguish a local experiment from a persistent, secure, supervised service.
The recording's use of Discord Administrator is preserved only as narrative evidence. Current public guidance uses the minimum permissions named by OpenClaw and Discord. The recording's permissive installer mode and suggestion to disable a timeout are not public recommendations.
OpenClaw is documented as a single trusted-operator personal-assistant boundary. Separate sessions do not create hostile multi-tenant isolation. Separate gateways and preferably separate OS users or hosts are required when trust boundaries differ.
Plaintext tokens can remain agent-readable even in local configuration and .env files. Public guidance uses SecretRefs where supported and pairs them with filesystem and tool restrictions.
The current OpenClaw security guide warns that smaller or heavily quantized models can be less resistant to prompt injection and tool misuse. This is an OpenClaw project position, not a benchmark performed by Venture Step.
The Local Model versus API Model page provides a reproducible test protocol. It does not claim that a production comparison was run. Current provider prices, terms, models, and data-handling rules require a date-specific check.
Claims held from publication
Do not publish the transcript's bot message-length estimate, search-service allowance, provider price estimate, model list, subscription detail, or setup-duration claim without a current primary source and a clear reader need.
Do not reproduce any token, API key, service-account value, phone number, email, user ID, server ID, channel ID, local path, gateway token, or private configuration from the recordings.
Do not say the local model was fully verified because it answered a direct prompt. Do not say Discord was fully configured because the bot appeared online. Do not say the dashboard worked.
Do not recommend Administrator, public bot distribution, open direct messages, wildcard allowlists, disabled timeouts, broad installer autonomy, tool access before channel verification, or operation on a primary personal or work machine.
Citation and refresh record
Public claims link to the source that controls them. OpenClaw releases, Node requirements, configuration keys, security checks, secrets, provider behavior, and Discord channel steps require a fresh check before publication.
Discord portal labels, scopes, permissions, intents, policies, and bot behavior require a fresh check. Gemma variants, license, context, runtime packaging, and provider pricing require a fresh check.