Build Your Own AI Assistant (Better Than OpenClaw) - No Subscription, No Tokens

Hi, this is Mike Barnes. Today I'm going

to talk to you about uh an alternative

to the very popular project called Open

Claw. Um this is something that uh I

built. I've been working uh installing

Open Claw and playing with it for uh two

weeks since it came out. And uh

basically I felt that the risks uh for

the system were just too expensive. Uh

as you can see on my slide here uh high

costs is one of the uh real big

problems. Uh I've been watching some of

the uh people on YouTube saying that

they're spending a couple hundred

dollars a day on uh token costs using

anthropic and there are some people

suggesting uh lower costs of doing it.

But uh these costs are pretty hard to

control because as the knowledge base

grows in the system, it continuously

uses more and more uh uh tokens and

these tokens uh charge up and you just

never know how much it's going to be.

And then of course we've all been

hearing about the privacy risk because

uh the system has virtual complete

control over your system and some people

are also buying separate computers like

uh the Mac minis to do it. the system is

sort of inflexible. There's a limited

amount of of customization and uh also

vendor lock lock in risks. Um so I tried

to use open claw with

um I was getting good results for a

while and then the system would break

and I continuously tried. I did go out

to outside models. The performance was

certainly much better, but uh I decided

that I was going to scrap uh openclaw

and build something myself. Uh and I was

going to you uh use uh GLM5, which is a

model uh that runs for free on on a

website called z.ai

to help me develop uh the code. So what

I wanted to do is to come up with a

self-improving system that offered a web

dashboard uh that protected privacy and

was free and open source with uh no

subscription required. So uh this is

probably underestimating what the cost

would be on a monthly basis. Uh but uh

you you at a minimum would be spending

between $15 and $100 to be able to use

uh the various APIs. This says J GBD4,

but obviously GPD4 isn't available

anymore. We'd be using, you know, 5.2 or

something like that. But after we set up

this system, uh there's no per token

charge, no subscription fees, and you

have unlimited queries forever.

And uh we're doing all the processing on

this one locally. Now um we're using

Alama as the base. So it is possible to

get uh tokens through the cloud. There's

a limited amount that resets every week

uh with Alama. And you can buy more

tokens. and Alama has some very capable

uh cloud-based tokens if you want to go

that way that would be far less

expensive uh than than using uh uh

Anthropics or OpenAI uh or even uh

Google. So the architecture basically is

is that we developed a web dashboard. We

have a llama runtime. uh the uh coding

which I've done for all the agents was

done on GLM5

and uh we have uh local fast models

running go out and query uh GLM5

on more advanced requirements I've built

in a document management system I've

created autolearning and knowledge

persistence so um here are some of the

foundations of how to set this up you

have to install NodeJS uh plus a lama

NodeJS is required to install Alama. You

want to pull the models that you're

going to use. Uh in this particular

case, I used OSS uh 20B. Uh I used uh

Kimmy 2.5 cloud, Lava and Gemma. Uh and

Lava and Jima are both u vision models

which would allow me to uh do OCR and

also identify um images.

When once we get our MPM start, we can

then use a browser uh and to use

localhost 3000 to to operate this

system. Uh we have a natural uh chat

interface through the browser as well as

through um I set this up to go with uh

telegraph.

I had been using WhatsApp with open call

but open call uh but when I used um

WhatsApp it interfered with my my work

because I do use it for work. Um I added

because of the work I do I work with the

federal government. I gave it the API

for SAM.gov. I put a capability to do

research there. I added a mo a model

selector. uh I created it so that uh it

improves itself. So I think that the

local process uh it wins because it's

cheaper, it's safer and better. And so

what I want to try to do here is to show

you what I built. Uh you can use this as

a reference uh in in doing your own

build. and uh you will wind up with

something I think will be much less

expensive, far more customized

um and much more secure than than using

open call. So I'm going to start out

right now showing you um uh the the

telegram. So, I'm I'm going to bring up

u the Telegram web interface.

>> And uh I set this up in this almost the

exact same way that you would do open

call, which winds up with my own bot, my

own bot here. I can now hit uh a slash

and hit help. And it gives me all my uh

capabilities here.

>> So, the model that I have running is

GPTOSS20B.

Uh it's uh roughly equivalent to the

original chat GPT, but it is running

locally. I have it set up so that I can

use uh Miniax M uh.2.5 on the cloud for

very complex tasks. I've used up all my

tokens uh for this week. So I have that

will reset next week. And for vision, I

have Lava uh 7B, but you can also use uh

one of the Gemma models as well. All of

these can run locally. And I have the

system set up. So that for different

tasks it will select different models I

have uh research for various uh topics

that I'm interested in uh space being

the industry that I work in uh AI uh my

my own hobby and then uh I can create a

list of documents uh and I can overwrite

or edit a document and then if I want

this the system to remember something I

can uh press remember I can get a status

of tokens. I can get a list of the uh

models that are available uh by typing

slashmodels

and I can also get a uh a status uh and

then while I'm in this area here I have

uh basically a chat uh capability.

So, uh, this is fine for when I'm, um,

outside of my home environment, but when

I'm in my home environment, I want to

have something that's a little bit, uh,

friendlier and a little bit easier to

use. So, um, what I did is I created

this this web system right here that you

see, and I called this, uh, Mike's

research assistant. Um, as you see here,

it's a very simple user interface. If

you've ever worked with chat GPT, uh

you'll you'll understand this. Uh I have

uh I believe my models right now. So

I'll just make sure that it's running uh

by uh typing hello and I should get some

sort of response if the uh Okay. So

hello Mike, can I support you? Uh

whether it's digging in trends for DAT,

which is a company I work for, scouting

new opportunities or just brainstorming

ideas. This is coming from the local

model. So I'm going to ask it to just

with a a button here to give me industry

news on space. So it's given here uh the

introduction and then I can uh get a

link. So uh if I uh press here, it'll

open up a new tab. The new tab will then

give me the article that it it's found

that uh uh it it believes that I would

be interested in. So, I can go back uh

to go right back to my local host. Now,

uh I have this set up here uh that it's

I've got the autolearn

and uh this means that as I'm

researching, I can I can uh learn uh

from these. Now, let's say I want to go

look at SAM.gov gov and I want to use

their API to see if there's any relevant

um opportunities for me to bid on. This

is going to go out. It's going to use

the keywords that I've given it and it's

going to give me a list of opportunities

here that I may want to uh to look at.

Um I can uh search on another topic here

which is uh AI and defense news. Um and

then uh I can get a a briefing on just

basically topics that I'm interested in

because these are things that I've been

looking at. Now in addition to this I've

uh what I I've decided to create a dream

team of agents and these uh dream teams

uh basically what they do is uh they

have certain uh capabilities or

pre-prompts that tell them how to do

certain things. So down here you'll see

the agent. I have a researcher, a

copywriter, a marketeteer, and a

strategist. And each one of them can be

assigned to a different LLM. They each

have their own prompt, and I can uh uh

ask them to do different things. So if I

click on uh the copywriter, for example,

uh I am now conversing with a different

personality than if I converse with the

researcher.

uh another capability. So uh these are

these here are my uh experts here. So I

can uh I have pre-built workflows for

proposal development uh doing

competition and uh uh and if I want to

go into doing uh a market entry, this is

all pre-prompted to let me to do that.

Um, I go back to the I can go back here

to the chat and uh I can uh have it do

with a uh slash research will do

research for me, but I have access to my

local model and I can have it do silly

things like u write a poem about space

debris

and just like chatg

it'll uh go off and and uh do whatever

task I want to. You can see the working

uh that that's going on right now. Um

as it's doing that uh let me show you

another capability I added here is let's

say I'm taking notes that I want it to

remember. I can uh come in here and say

uh please

uh you know please uh remember I have a

meeting at 3 p.m. on the 21st

and it will then uh add that to uh my

memory. Um, I also have the ability to

uh create documents that I can store.

And one of the other uh capabilities

I've had is to attach. So, this attach

will allow me to add a document or it

will allow me to um

uh to add an image. And if I'm using

let's say uh one of the vision models

like lava, I can then have that document

uh broken down in uh you know or OCR

scanned or I can have an object uh

described by attaching the op uh the

object.

>> So I built this uh using u uh basically

uh GLM uh 5.0.

Um, anyone really can outline their own

workflow uh to build this. Uh, uh, this

once again is running on my own, uh,

computer. And, um, let's see here. So,

here's here's the poem that was uh, uh,

written, you know, uh, uh, silent shr

uh, shreds of yesterday's ambition.

Titanium ghosts of Newton's dance. A

screwdriver spins through the endless

midnight where once the star is held a

watchful trance. So it's sort of amazing

that without spending any money on

tokens without uh paying uh for any

subscriptions I'm able to build this

system and this system uh with the

capabilities of the smart team I can add

other capabilities to this as well. So,

I offer this up as sort of a model that

can be used uh for people that don't

want all the limitations. Once again,

nothing is going outside of my own

environment. Uh this is running

completely on my computer using my own

uh RTX 4080 processor. I have a 5090 on

uh my laptop.

I can access this from outside of my own

premise. I offer this up as an

alternative to open claw. Uh leave it to

people who are more inclined to do

programming than I am uh to build upon

this idea. But I think this is a better

model than open claw.

Although I have to uh to give credit for

openclaw for kind of uh instigating me

into doing this project. And also want

to thank Randy Hill who um is my friend

and he's also uh the chief technology

officer for Govotics because many of

these concepts uh he developed on an

enterprise scale months ago on a a

project he called Japetto. So um this is

uh a home version uh but I think it's a

it's a very good model uh that can be

used uh for anyone who wants to build a

secure lowcost

uh assistant.