Stop Paying for Claude Code (3 Ways to actually use it for FREE!!) — Transcript
Full transcript
- 0:00Claude Code is $20 a month, but not
- 0:03really. $20 [music] is what the AI
- 0:05inside it costs. Claude Code itself, the
- 0:08app, the tool, the thing that actually
- 0:10writes your code, is free. You just need
- 0:13to put a free engine inside it. Today,
- 0:15I'll show you three ways to do exactly
- 0:18that. [music] Let's build. Quick picture
- 0:20before we start. Think of a sports car.
- 0:22There's the body, the steering wheel,
- 0:24the dashboard, everything you interact
- 0:26with. And then, there's the [music]
- 0:28engine. Claude Code is the body. The AI
- 0:31is the engine. Today, we swap the engine
- 0:33three times. One runs [music] free on
- 0:35your laptop, one runs free in the cloud,
- 0:38and the third automatically switches
- 0:40between free AI models when you hit a
- 0:42limit.
- 0:47Okay, let's get into it. The first
- 0:50engine we're plugging in is called
- 0:51Ollama, and it's completely free. Just
- 0:55head over to ollama.com and hit
- 0:57download. It works on Mac, Linux, and
- 0:59Windows. I'm so I'll select Windows.
- 1:03Now, you can install it using a terminal
- 1:05command, or just download the installer.
- 1:07I'm going with the installer. Once it's
- 1:09downloaded, double-click the setup file
- 1:11and follow the installation steps. And
- 1:14that's it. Ollama is ready to use. Click
- 1:17the icon in the top left. Three things
- 1:19matter here. First, apps. Coding tools
- 1:22like Claude Code, Codex, Hermes, and
- 1:24others. Now, clicking one doesn't
- 1:26install anything. It just copies the
- 1:29command you need to connect Ollama to
- 1:30that tool. We'll use Claude Code in a
- 1:32minute. Second, chat. This is where we
- 1:35test models. Third, settings. The cloud
- 1:39toggle shows models on Ollama's servers.
- 1:42Those run on Ollama's servers. We're
- 1:44focusing on models that run directly on
- 1:46our laptop, so we can leave this off.
- 1:48Now, the one that actually matters,
- 1:51context length. Context is how much of
- 1:53your project the model can keep in its
- 1:55head at once. Claude Code works by
- 1:57reading your files, understanding them,
- 2:00then editing them. All of that has to
- 2:02fit inside the model's context. If it's
- 2:05too low, the model can lose track of
- 2:07your project as it works. Set it to 64K.
- 2:10That's the fix. Open chat and pick Qwen
- 2:123.54B,
- 2:14the small one, just to warm up the
- 2:15system. Type hi. First message always
- 2:18downloads the model, so it takes a
- 2:20second. After that, the model is already
- 2:23downloaded, so you won't have to
- 2:24download it again. But, hi doesn't prove
- 2:27anything. Time for the real test. Now,
- 2:29watch this. Wi-Fi off. Browser's dead.
- 2:33No internet on this machine at all. And
- 2:35I ask it to make a simple calculator
- 2:37app. And it writes it offline. And you
- 2:40don't have to worry about hitting a
- 2:42usage limit or waiting 5 hours,
- 2:44sometimes even a week, for it to reset.
- 2:46You can keep using it as much as your
- 2:48hardware allows. Every word of that code
- 2:51came from a model sitting on this
- 2:53laptop. And here's the biggest
- 2:55advantage, privacy. Your code and
- 2:58prompts stay on your computer. I'm
- 3:00switching the Wi-Fi back on for one
- 3:02reason, to paste the code into a free
- 3:04online compiler. And it runs. Is it
- 3:07pretty? No. But, a laptop with no
- 3:09internet just wrote working software. No
- 3:12subscription, no usage limits, nobody
- 3:15can switch it off. That's Engine 1. But,
- 3:17which model should you actually run?
- 3:20Don't memorize charts. Ask an AI. First,
- 3:23check your hardware. Press control plus
- 3:25shift plus escape, then click the
- 3:27performance tab. Mine says 12 GB of
- 3:29VRAM, that's your graphics card's
- 3:31memory, and 32 GB of RAM. Now, paste
- 3:34this into ChatGPT or any AI you have
- 3:37open. I have 12 GB of VRAM and 32 GB of
- 3:41RAM. I want to run a local model on
- 3:43Ollama for coding. Select the best model
- 3:45for my use case and reply with its exact
- 3:48Ollama name. It said Qwen 2.5 Coder 14B.
- 3:52Decent answer, but always ask a
- 3:54follow-up because the first answer might
- 3:56be outdated and new models keep coming
- 3:59out. Is this still the absolute best
- 4:01choice as of the current model releases.
- 4:03Check the latest available models before
- 4:05answering. And it corrected itself. Qwen
- 4:083.59B.
- 4:10That one extra question could save you
- 4:12from running a model that's already been
- 4:14replaced by something better. Back to
- 4:16Ollama to grab Qwen 3.59B.
- 4:18And here's the first problem. It's not
- 4:21in the drop down. Ollama only shows a
- 4:23selection of popular models in that
- 4:25menu. The model exists. It's sitting in
- 4:27Ollama's library, but to install it
- 4:30you'd have to find the exact command,
- 4:32open a terminal, and run it yourself.
- 4:34And that's where a simple setup can turn
- 4:36into an hour of Googling, Stack Overflow
- 4:39tabs, and random GitHub issues from
- 4:412023. But what if you didn't have to
- 4:43figure out any of that yourself? That's
- 4:46where another AI comes in. The one I'm
- 4:48using is called Anti-Gravity, a free
- 4:51coding tool from Google. What makes it
- 4:53different from a normal chatbot is that
- 4:55it can actually work on your computer.
- 4:57It opens your terminal, runs commands,
- 5:00reads error messages, and even fixes
- 5:02problems along the way. So you tell it
- 5:04in plain English what you want
- 5:05installed, and it just does it. To use
- 5:07it, download Anti-Gravity, sign in with
- 5:09a Google account, open it up, and give
- 5:11it one sentence. I am running AI models
- 5:14locally. Install the model Qwen 3.59B
- 5:18and set up everything so it works with
- 5:20Ollama. It finds the model, downloads
- 5:22it, hooks it into Ollama, done. One
- 5:25thing to watch. Anti-Gravity will ask
- 5:27for permission a few times while it
- 5:29works because it's running commands on
- 5:31your machine. Read each prompt before
- 5:33you click allow. You're the driver. It's
- 5:35the mechanic. When it's finished, close
- 5:38Ollama and reopen it. The new model
- 5:40shows up in the drop down. And this is
- 5:42the house rule for the whole video.
- 5:44Anything longer than one line, you don't
- 5:46type it yourself. An AI installs it. If
- 5:49something breaks on the way, copy the
- 5:51error, paste it back into the chat, and
- 5:53say, "Fix this." Now for the part you
- 5:55came here for. Back to Ollama, click the
- 5:58apps tab, and click Claude code. See
- 6:00that little flash in the corner? That's
- 6:02the install command for Claude code
- 6:04copied to your clipboard. Make a folder
- 6:06anywhere on your computer. Mine is
- 6:08called free Claude code. Open it, right
- 6:11click in the empty space, choose open in
- 6:13terminal, paste, and hit enter. If it'll
- 6:16say Claude code is not installed,
- 6:18install now? Type Y. And this is
- 6:20important. What's installing right now
- 6:23is the real terminal-based Claude code,
- 6:25Anthropic's genuine app, straight from
- 6:27its official source. Not a copy, not a
- 6:30clone, not a wrapper. Ollama just runs
- 6:32the official installer for you. When it
- 6:34finishes, it asks which model to use.
- 6:37Pick Claude 3.59B.
- 6:39And there it is, Claude code sitting in
- 6:41your terminal with a free engine inside
- 6:44it. No Anthropic subscription, no API
- 6:46key. You're running it with a local
- 6:48model through Ollama. Let's put it to
- 6:50work. Type, "Design me a simple
- 6:53website." Or, if you want to keep it
- 6:55even simpler, "Build me a calculator
- 6:57app." It'll ask you a question or two to
- 6:59understand what you want, then ask
- 7:01permission to create files in the
- 7:02folder. Allow it. First draft is going
- 7:05to be rough, honestly. But, that's the
- 7:07whole point. You iterate. Make the
- 7:09buttons bigger, switch to dark mode,
- 7:11clean up the layout. Each round, it gets
- 7:13closer. And remember, this is all
- 7:15running locally, with no cloud AI usage,
- 7:18no provider imposed usage limits, and it
- 7:21still works with the Wi-Fi off. Next
- 7:23time you want to use it, you don't
- 7:24repeat any of this setup. Right click
- 7:26your folder, open in terminal, and type
- 7:29one line. That's the whole routine from
- 7:31now on. So, why isn't this video over?
- 7:34Because you saw it, too. It's slow. And
- 7:36honestly, it's not that smart. So, let's
- 7:39upgrade. The next engine runs in the
- 7:41cloud. Faster, smarter, still free. And
- 7:44this is where the model menu starts
- 7:46lying to you. The second engine is
- 7:48called OpenRouter. Think of it as a
- 7:50middleman that gives you access to
- 7:52dozens of AI models through one single
- 7:54key. And a bunch of those
- 7:56[clears throat] models are completely
- 7:57free. Nothing runs on your laptop this
- 7:59time. The AI lives on their servers. You
- 8:02just need one key to talk to it. Head to
- 8:05openrouter.ai and sign up. Free, no
- 8:08credit card. Once you're in, click API
- 8:10keys, then create key. Give it a name.
- 8:13Mine is Claude code, and set an
- 8:15expiration date. I set mine to 7 days,
- 8:18which means by the time you're watching
- 8:20this, my key is already dead. If you
- 8:22ever add money to this account later,
- 8:24set a spending limit, too. It's a
- 8:26seatbelt, wear it. Create the key and
- 8:29copy it somewhere safe right now,
- 8:31because Openrouter only shows it to you
- 8:33once. And treat it like a password.
- 8:36Anyone who gets this key can burn
- 8:37through your quota. Before we plug this
- 8:39key into anything, look at what you just
- 8:42unlocked. Click models at the top, open
- 8:44the price filter, and set it to free.
- 8:47Look for models with the colon free
- 8:49label. Those are the ones you can use
- 8:50without buying credits. Every request to
- 8:53those models costs zero. Now, here's the
- 8:55catch. This list rotates. Models join,
- 8:59models leave, companies pull their free
- 9:01tiers all the time. So, the names you
- 9:03see today might not be the names you see
- 9:05next month. The names don't matter. The
- 9:08filter does. As long as you know how to
- 9:10find the free list, you'll always have
- 9:12an engine. Now, we plug this key into
- 9:14Claude code. The block of code you need
- 9:16is sitting in the description of this
- 9:18video. Copy it. You only change one
- 9:20thing in it. Your key goes between these
- 9:22two quotes. The rest of the block is
- 9:25just instructions that tell this
- 9:26specific terminal window to send all its
- 9:29work to Openrouter instead of Anthropic.
- 9:31One of those lines clears out the old
- 9:33key slot, so there's zero chance of
- 9:35accidentally falling back to a paid key.
- 9:38And these three lines near the bottom,
- 9:39the ones labeled Opus, Sonnet, and
- 9:41Haiku, I added those myself on purpose.
- 9:45Here's why. Claude code's model menu has
- 9:47three seats inside it, named after
- 9:49Anthropic's paid models. These lines map
- 9:52each of those seats to a free model on
- 9:54Openrouter. So, whichever seat you pick
- 9:56inside Claude code, the model we assign
- 9:59to that seat is the one answering. Hold
- 10:01on to that. It matters in a minute. Copy
- 10:03the whole block. Back to the folder we
- 10:05made earlier, right click, open in
- 10:08terminal, paste, hit enter. Claude code
- 10:11opens up again. Same app you were using
- 10:135 minutes ago, completely different
- 10:15engine behind it. Say hi. The reply
- 10:17comes back before you finish blinking.
- 10:19That's the difference between a model
- 10:21running on your laptop and a model
- 10:23running in the cloud. And if you want
- 10:25proof of where the work is actually
- 10:26going, type {slash} status. See that
- 10:29address? That's OpenRouter. That's where
- 10:31every message you send from now on is
- 10:33traveling. Now, remember the model menu
- 10:36I teased earlier? Type {slash} model.
- 10:38Look at that. Sonnet, Haiku. Pick
- 10:41Sonnet. Ask it literally anything. It
- 10:44answers. So, did we just get Claude
- 10:47Sonnet for free? Honestly, no. And I'll
- 10:50prove it in one screenshot. Open
- 10:52openrouter.ai/activity
- 10:55in your browser. This is the real log.
- 10:57Every request you've sent, the exact
- 11:00model that actually answered it, and the
- 11:02price. All zeros. And nowhere on this
- 11:05page does the word Claude appear. The
- 11:07menu inside Claude code shows seat
- 11:09names. OpenRouter's activity log shows
- 11:11the actual passenger. Any video out
- 11:14there calling this free Sonnet is
- 11:16reading the seat label and not checking
- 11:18who's actually in the seat. One more
- 11:20thing that might scare you if you don't
- 11:22know about it. Open Claude code's own
- 11:24usage screen and it'll tell you this
- 11:26session cost you around 45 cents. Relax.
- 11:30That number may be an estimate based on
- 11:32Anthropic's pricing, not what OpenRouter
- 11:34actually charged you. So, don't judge
- 11:36the bill by that screen. The number
- 11:38that's actually real is the one on
- 11:40OpenRouter. And on OpenRouter, you paid
- 11:43zero. But, there's a trade-off. Unlike
- 11:46the local setup, your prompts and
- 11:48project files are now being sent to
- 11:50external servers. So, you need to be
- 11:52careful about what you share. And you're
- 11:54also dependent on their servers, their
- 11:56free models, and whatever usage limits
- 11:59they impose. Okay, so the same task that
- 12:01took the local engine a few minutes, the
- 12:03cloud writes it in seconds. But, here's
- 12:06the honest answer. OpenRouter's free
- 12:09tier is capped. For models marked
- 12:11{colon} free, you're limited to 20
- 12:13requests a minute and 50 requests a day
- 12:16on a new account.
- 12:17>> 50 sounds like plenty until you remember
- 12:19what Claude code actually does. One task
- 12:22from you can trigger 10, 15, 20 requests
- 12:25behind the scenes because it's reading
- 12:27files, writing files, checking its own
- 12:29work, then fixing what it broke. One
- 12:32real coding session can empty your whole
- 12:34daily cap. And [snorts] when you hit
- 12:36zero, you get error 429, rate limited,
- 12:39mid task, frozen, come back tomorrow.
- 12:42That wall right there is the exact
- 12:44reason the third method exists. The
- 12:46third setup is called Omni route. Here's
- 12:49what it actually does. It's a free app
- 12:51that connects Claude code to over 375
- 12:54providers and gives you access to more
- 12:57than 1,400
- 12:58AI models, including Claude, GPT,
- 13:01Gemini, Kimi, and Deep Seek. Open router
- 13:04has a free cap. Fine, but there are
- 13:06other services out there with their own
- 13:08free caps, Nvidia, Hero, Kline, a bunch
- 13:12of others. On their own, each one is
- 13:14small. Together, they're huge. And
- 13:17that's where Omni route comes in.
- 13:19>> [clears throat]
- 13:19>> It brings all of them together in one
- 13:21place. When one provider runs out of
- 13:24free usage, it can switch to another one
- 13:26automatically. You don't have to keep
- 13:28changing providers yourself. One AI runs
- 13:30out of fuel, another one takes over. And
- 13:33that's what we're setting up today.
- 13:34We're going to install it the same way
- 13:36we fixed Ollama earlier. Anti-gravity
- 13:38does the work. Open Anti-gravity, paste
- 13:41this prompt from the description, and
- 13:43let it go. It'll install Claude code if
- 13:45you don't have it already, install Omni
- 13:47route, start the service, and open the
- 13:49dashboard in your browser. While it's
- 13:51working, it'll ask for permission more
- 13:53than three times. Read each one before
- 13:56you click allow. It's running real
- 13:58commands on your machine. Once the
- 13:59installation finishes, click the
- 14:01localhost link in the dashboard. Then
- 14:04the little arrow at the top right opens
- 14:06it in your browser. First time you log
- 14:08in, the default password is just the
- 14:10word change me. Paste it in and change
- 14:12it to something real in settings right
- 14:14after. All right, now the fun part.
- 14:17We're going to start connecting engines
- 14:18to this thing. On the dashboard, open
- 14:21the Omni proxy tab. This is where you
- 14:23plug in every free service you want Omni
- 14:25route to pull from. We'll do five of
- 14:27them. First up, Kiro. Click the
- 14:29providers option, search Kiro, accept
- 14:32the warning that pops up, and sign in
- 14:34with your free Kiro account. Done. Now
- 14:36look at its model list. Yes, those are
- 14:39real Claude models. Sonnet, Haiku, the
- 14:42actual ones. But here's something you
- 14:44need to know before you get too excited.
- 14:46Kiro's free tier is 50 credits a month,
- 14:49which is maybe a handful of prompts
- 14:50before it's gone. And Omni route itself
- 14:53puts a warning on the screen saying
- 14:55Kiro's terms of service don't
- 14:56technically allow this kind of use. So
- 14:59use it at your own risk. I'm showing you
- 15:01how it works, but whether you want to
- 15:03connect your account is entirely your
- 15:05call. Second, the two that take a key.
- 15:08Nvidia [clears throat]
- 15:09is the backbone of this setup. Create a
- 15:11free API key at build. nvidia.com. Paste
- 15:14it into the Nvidia card, click check,
- 15:16and set its priority to one so it's used
- 15:19first. Nvidia mixes free and paid
- 15:21models, so don't worry about
- 15:23accidentally using a paid one just yet.
- 15:25We're going to filter those out in a
- 15:26minute. Then open router. Paste the same
- 15:29free key from method two. Same free
- 15:31models, same limit, 50 requests a day.
- 15:34Next, the two that just need a sign in.
- 15:37Klein, an AI coding assistant whose new
- 15:39models are often free for a while. Sign
- 15:41in with a free Klein account. And
- 15:43anti-gravity. Sign in with Google to
- 15:45unlock Gemini and some Claude models
- 15:47within a usage limit. Same warning as
- 15:50Kiro though. Google's terms don't allow
- 15:52this, and Omni route flags it right
- 15:54here. Your account, your call. Back on
- 15:56the home dashboard, mine shows four
- 15:58connected. Klein threw an error, but
- 16:00four is plenty. And stay with me because
- 16:02in about two minutes, we're going to get
- 16:04the real Claude to answer through this
- 16:06exact setup. Now we put all of this into
- 16:09one combo. A combo is just Omni route's
- 16:12word for a pool of engines that work
- 16:14together. Click the combos tab, hit
- 16:16create your first combo, and give it a
- 16:18name. Mine is Claude. Under steps, you
- 16:21pick the first provider, the model to
- 16:23use, and the account. I'm starting with
- 16:26GPT-OSS-120B
- 16:28on Nvidia. Click add. Then Omni Route
- 16:30asks you how it should pick between all
- 16:32your engines when a request comes in.
- 16:34There are multiple strategies. I'm going
- 16:37with intelligent auto. It makes the
- 16:39decision for me every single time, which
- 16:41is exactly what we want. The whole point
- 16:43is to stop manually switching between
- 16:45providers. Alternatively, you could also
- 16:48choose round robin or any other you
- 16:50like. Next, in the candidate pool below,
- 16:52pick every provider you connected. Back
- 16:55in steps, add a second engine, Open
- 16:57Router this time as the backup. The up
- 17:00and down arrows next to each engine set
- 17:02the order they get tried in. Top one
- 17:04goes first. Hit create. Hit save. And
- 17:07just like that, you've got your engine
- 17:09pool ready to go. One warning before we
- 17:11move on. When you add Open Router as
- 17:14that second engine, it'll ask you which
- 17:16specific model to use from Open Router.
- 17:18And one of the options in that drop down
- 17:20is also called auto. Normally, you don't
- 17:23want to pick that one. Open Router's
- 17:26auto can quietly route your request to a
- 17:28paid model and charge you. But for now,
- 17:30just pick it. We'll fix that in a
- 17:32moment, so you don't have to worry about
- 17:34accidentally getting charged. Now we
- 17:36need to connect Claude Code to this
- 17:38engine pool we just built. For that,
- 17:40Omni Route gives you its own API key.
- 17:42Think of it as the master key that
- 17:44unlocks every engine inside the combo.
- 17:47Head to the API keys tab inside Omni
- 17:49Route, hit create, copy the key, and
- 17:51save it somewhere safe. Now flip back to
- 17:54Anti-Gravity one more time. There's a
- 17:56prompt in the description called
- 17:57shortcuts. Paste it in. This prompt does
- 18:00three things at once. It takes your Omni
- 18:02Route key and saves it into a config
- 18:04file. It creates a dedicated work folder
- 18:07on your computer, and it drops two
- 18:09shortcuts right on your desktop. One
- 18:11that opens Claude Code already wired up
- 18:13to Omni Route, and one that opens the
- 18:15Omni Route dashboard in your browser.
- 18:17Halfway through the install,
- 18:19Anti-Gravity will stop and ask you for
- 18:21the key. Important, don't paste it into
- 18:23the chat. That's the fastest way for a
- 18:25key to leak. Instead, it'll open a file
- 18:28for you. Paste the key into that file,
- 18:30save it, and type continue in the chat.
- 18:33If the file doesn't open on its own,
- 18:35just say open notepad so I can paste the
- 18:37key. Done. Now check your desktop. Free
- 18:40Claude code and Omni route dashboard.
- 18:42Two icons ready to go. But before you
- 18:44double click anything, there's one more
- 18:46trap you need to know about. Each
- 18:48service we connected, Hero, Nvidia,
- 18:51OpenRouter, all of them, has both free
- 18:53and paid models on their platform. And
- 18:56when Claude code's model menu opens up,
- 18:58it shows you all of them mixed together
- 19:00with no label telling you which is
- 19:02which. So you could accidentally pick a
- 19:04paid model, send one message, and end up
- 19:06with a bill or an error message if you
- 19:08have no credits. To fix it, there's one
- 19:11more prompt in the description. I called
- 19:13it the free models prompt. Paste it into
- 19:15anti-gravity. It goes through every
- 19:17connected provider, filters out every
- 19:19paid model from the menu, leaves only
- 19:22the free ones visible, and runs a test
- 19:24to confirm the fix held. From that point
- 19:26on, the models you see in that menu are
- 19:28the ones we've configured as free. No
- 19:30more guessing. All right, moment of
- 19:33truth. Double click the free Claude code
- 19:35shortcut on your desktop. Terminal
- 19:37opens, Claude code loads, say hi. It
- 19:40replies. Now ask it what model it is,
- 19:43and it tells you Claude. But you know
- 19:45the rule from method two, a model's word
- 19:48isn't proof. Some free models will even
- 19:50claim to be Claude. So check the
- 19:52receipt. Open the OmniRouter dashboard,
- 19:55go to logs, and look at the last
- 19:56request. Provider, Hero. Model, Claude
- 20:00Sonnet. That's the real Claude Sonnet
- 20:02talking to you right now, routed through
- 20:04Hero's free credits. Hero's free credits
- 20:07are tiny, so this won't last long. When
- 20:09they run out, switch to the Pombo we
- 20:11built. Inside it, Nvidia and OpenRouter
- 20:14take over on their own. Just know that
- 20:16from then on, it isn't Claude answering.
- 20:18It's the free engines. You keep working,
- 20:20and the switch happens behind the
- 20:22scenes, and the bill stays at zero. All
- 20:25right, let's put this whole setup
- 20:26through a real test, not a toy prompt,
- 20:29something you'd actually want to build.
- 20:31I typed in, "Design me a simple personal
- 20:33note-taking application." And watch what
- 20:36happens here, because this is the part
- 20:38most tutorials skip. Claude Code doesn't
- 20:40just start generating random code. It
- 20:42stops and asks me three questions first.
- 20:45Do I want a desktop app or a web app? Do
- 20:47I want my notes stored locally on my
- 20:49machine or in the cloud? And which
- 20:51features matter most to me? Tags,
- 20:54search, markdown, dark mode? I answer
- 20:57each one, it builds a plan based on my
- 20:59answers, shows me the plan, and only
- 21:01then starts coding. A few minutes later,
- 21:04I've got a working note-taking app
- 21:05running on my own computer on localhost.
- 21:08Open it in a browser, create a note,
- 21:10save it, close the app, reopen it, note
- 21:13still there. Fully working, zero dollars
- 21:16spent. Now, I want to be straight with
- 21:18you about something. If you've searched
- 21:20for free Claude videos before, you've
- 21:22probably seen a hundred titles screaming
- 21:25"Unlimited free Claude." It's not true.
- 21:28There is no setup on the internet right
- 21:30now that gives you unlimited access to
- 21:32the real Claude model for free. What you
- 21:34actually get with this build is two
- 21:36things stacked together. A small taste
- 21:38of real Claude through Keiro's monthly
- 21:40credits, plus the Claude Code tool
- 21:43itself running completely free with
- 21:45other engines inside it. That's the
- 21:47honest version. And here's the thing,
- 21:49for most people that stack is already
- 21:52enough. You get real Claude for the hard
- 21:54problems where you need its brain, and
- 21:56you get free open-source engines for
- 21:58everything else. You don't have to stop
- 22:00working just because one provider runs
- 22:02out of credits. You never see a bill as
- 22:05long as you stick to the free models and
- 22:07usage limits we've configured. And
- 22:09that's the real advantage of this setup.
- 22:11Not unlimited Claude, but a free coding
- 22:14workflow that keeps going even when one
- 22:16engine runs out. One bonus before we
- 22:18wrap up. If you don't specifically need
- 22:21Claude Code itself and just want to code
- 22:23every day with the best free AI models
- 22:25possible, there's a simpler setup. And
- 22:28honestly, it's the one I'd keep coming
- 22:30back to. It's called Kilocode and it
- 22:32lives inside VS Code, Microsoft's free
- 22:35code editor. One honest note before we
- 22:37start. This isn't Claude Code. It's a
- 22:40different coding agent made by a
- 22:42different team. But the idea is exactly
- 22:45the same as everything we just built. A
- 22:47free engine inside a coding tool. Think
- 22:49of this as the lightweight daily driver.
- 22:52Claude Code when you want the real deal,
- 22:54Kilocode when you just want to ship
- 22:55stuff without thinking about
- 22:57subscriptions. All right, let's build
- 22:59this one from scratch. Head to
- 23:01code.visualstudio.com
- 23:03and download VS Code. Install it, open
- 23:05it up, and the first screen asks you to
- 23:07pick a theme. I went with dark, but pick
- 23:10whatever you want. Once you're in, click
- 23:12open folder at the top left, choose any
- 23:15folder on your computer where you want
- 23:16to work, and when it asks if you trust
- 23:18the files inside, say yes. Now we add
- 23:21Kilocode. On the left side bar, click
- 23:23the extensions icon. It looks like four
- 23:26little squares. Search Kilocode, click
- 23:28install, and give it a second. Once it's
- 23:31installed, a new icon appears in that
- 23:33same side bar. Click it. Kilo asks you
- 23:36to sign in. Do that and you're done
- 23:38setting up. That's it. You now have a
- 23:40coding agent running inside VS Code and
- 23:43you haven't paid for anything. Open
- 23:45Kilo's model menu for a second. Right
- 23:47out of the box, it already comes with
- 23:49free models preloaded, Auto Free, Nvidia
- 23:52Neumotron, and a few others. So
- 23:54technically, you could start coding
- 23:56right now without connecting anything
- 23:58else. But this is Kilo's own curated
- 24:01free tier and the lineup changes every
- 24:03few weeks as deals come and go. So
- 24:05instead of relying only on what Kilo
- 24:07gives you, we're going to add our own
- 24:09providers underneath. That way, you've
- 24:12always got multiple engines on tap. And
- 24:14when one hits its limit, you can switch
- 24:16to another. Let's start with Nvidia, the
- 24:18same provider we used in Omni route. Go
- 24:21to Kilo settings, click providers, then
- 24:23custom providers, and hit connect.
- 24:26There's a block of connection details in
- 24:27the description. Paste it in. Then, grab
- 24:30your Nvidia API key, the same one you
- 24:32created earlier or a new one if you
- 24:34prefer, and paste it into the key field.
- 24:37When you hit submit, Killo pulls down
- 24:39the full list of every model Nvidia
- 24:41offers. It's long, and you don't need
- 24:44all of them. We're just going to pick
- 24:46three. Click deselect all at the top,
- 24:48then individually tick these three:
- 24:50Nemotron 3 Ultra, GLM 5.3, and GLM 5.3
- 24:55Flash. These are the three free models
- 24:58we're using for this setup. Below the
- 25:00list, turn on reasoning so the models
- 25:02can think through longer problems, and
- 25:04turn on images only for models that can
- 25:07actually read images. If you flip it on
- 25:09for a text-only model, it'll error out.
- 25:11Hit submit. Nvidia is now live inside
- 25:14Killo, and honestly, Nvidia is my
- 25:17favorite provider in this whole video.
- 25:19Its free tier is generous, with a limit
- 25:21of around 40 requests a minute according
- 25:23to the current offering. That's plenty
- 25:25for testing and personal projects, which
- 25:27is exactly what we're doing. All right,
- 25:29let's test it with something real. I
- 25:32opened Killo, picked GLM 5.3 Flash from
- 25:35the model drop-down, the fastest of the
- 25:37three Nvidia models we just added, and
- 25:39gave it one prompt. It's in the
- 25:41description. Just a short list of what I
- 25:43wanted: a start screen, enemy waves,
- 25:45power-ups, a boss, and a high score. And
- 25:48watch what happens. It wrote the entire
- 25:51game. One file, around 1,200 lines of
- 25:54code. Double-click it, and the game runs
- 25:56in your browser. Start screen with
- 25:58instructions, enemy waves that get
- 26:00harder as you go, power-ups that drop
- 26:02from enemies, and a boss fight every
- 26:05fifth wave. And this isn't just a
- 26:07screenshot of a game. I can actually
- 26:09play it. I played it for about 10
- 26:11minutes while editing this video, and
- 26:12the whole thing, concept to playable,
- 26:15cost me zero dollars. That's the kind of
- 26:17thing you can start doing with these
- 26:19free models today. Now, that free Nvidia
- 26:22tier is generous, but it's still a tier.
- 26:25You will hit the limit eventually,
- 26:27especially on bigger projects. And this
- 26:29is where the setup gets useful. Instead
- 26:32of waiting for your limit to reset, you
- 26:34just add another engine. The setup is
- 26:36the same every time. Open settings,
- 26:38providers, custom providers, connect,
- 26:41and repeat the exact same flow we just
- 26:43did for Nvidia, but with different
- 26:45services. First, open router. Paste in
- 26:48the same key you created earlier, and
- 26:50this time you've got access to its free
- 26:52models like Nemotron 3 Ultra, Inkling,
- 26:55and Gemma 4. Limit, 50 requests a day.
- 26:58Next, Gemini. Head to Google AI Studio,
- 27:01generate a free key, and paste it in.
- 27:03The connection details are in the
- 27:05description. One important thing to flag
- 27:07about Gemini, on the free tier
- 27:09specifically, Google may use your
- 27:11prompts to improve its products. So, if
- 27:13you're writing anything sensitive,
- 27:15client code, proprietary logic, anything
- 27:17you don't want being used for model
- 27:19improvement, don't route it through
- 27:21Gemini's free tier. Use it for personal
- 27:23projects instead. And last one, Kline.
- 27:26This one isn't added inside Kilo. It's
- 27:28its own extension, and you don't need an
- 27:30API key. Install the Kline extension
- 27:32inside VS Code, sign in with their
- 27:34account, and you're set. The interesting
- 27:37thing about Kline is that it
- 27:38periodically changes which free models
- 27:40it offers. So, the lineup you see today
- 27:42might be different next week. And there
- 27:44are plenty more free providers out
- 27:46there. I've listed them all in the
- 27:48description. Now, here's a question
- 27:50you're probably going to run into
- 27:51eventually. What happens when you're
- 27:53mid-task, mid-code generation, and the
- 27:56model you're using suddenly hits its
- 27:58daily limit? Kilo handles this pretty
- 28:00keenly. Inside the same chat thread,
- 28:03just open the model drop-down at the
- 28:04bottom and pick a different model. Kilo
- 28:07carries forward the available
- 28:08conversation and project context, so the
- 28:11new model can pick up where you left
- 28:12off. You don't have to start explaining
- 28:14the entire project from scratch. And if
- 28:17you're starting a brand new chat and you
- 28:18want Kilo to remember your project long
- 28:20term, what you're building, what's done,
- 28:23and what's next, turn on its memory
- 28:25feature by typing {slash} memory on.
- 28:27From that point on, Kilo keeps running
- 28:29notes about your project across
- 28:31conversations. So, when you switch
- 28:33models or start a new chat, that saved
- 28:35project context can help you get back to
- 28:38work without starting from zero. So,
- 28:40we've built four completely different
- 28:42setups. And the real question is, which
- 28:44one should you actually use day-to-day?
- 28:47Here's the honest answer. If you've got
- 28:49a decent graphics card and privacy
- 28:51matters to you, especially if you're
- 28:53working with code you don't want sitting
- 28:55on someone else's server, go with
- 28:57Ollama. It works offline, but you're
- 28:59limited by your own hardware. If you've
- 29:01got no graphics card and just want speed
- 29:03today with no hardware investment,
- 29:06OpenRouter is your pick. Just remember
- 29:08the 50 requests a day cap on the free
- 29:10models. If you want the real Claude code
- 29:12experience with multiple free engines
- 29:14behind it, Omni route is the one to try.
- 29:17And if you don't specifically need the
- 29:19Claude code interface and just want a
- 29:21straightforward way to use free coding
- 29:23models, Kilo code inside VS Code is
- 29:26probably the easiest daily driver. And
- 29:28whichever one you choose, there are four
- 29:30things you need to know before you start
- 29:32using it seriously. Here they are.
- 29:34One, nothing is unlimited. Every free
- 29:37tier has a cap, and real Claude is never
- 29:40free for long. The whole point of this
- 29:42setup is to combine those limits, not
- 29:44pretend they don't exist. Two, free
- 29:47cloud engines see what you send. Client
- 29:50code, passwords, private files, anything
- 29:52sensitive, keep those on Ollama. Free
- 29:55doesn't automatically mean private.
- 29:58Always check what a provider does with
- 29:59your data before sending anything
- 30:01sensitive. Three, some connections may
- 30:04violate the provider's terms. Kiro and
- 30:06Anti-Gravity are flagged right on
- 30:08screen. Your account, your call. I'm
- 30:11showing you how these connections work,
- 30:13but you should always check the terms
- 30:14yourself before using them. Four, expect
- 30:17errors. We hit plenty. Copy the error,
- 30:21paste it into Anti-Gravity and say, "Fix
- 30:23this." That's part of working with AI
- 30:25coding tools. Sometimes the first
- 30:27attempt fails, but you can usually use
- 30:29the error message to figure out what
- 30:31went wrong. And one final rule, whatever
- 30:34a menu tells you, check the receipt, the
- 30:36actual usage and billing page, not the
- 30:39label. Everyone told you Cloud Code
- 30:41costs $20 a month. Today, we got the
- 30:44real Cloud Code, Anthropic's actual
- 30:47coding app running with free engines
- 30:49inside it three different ways. One
- 30:52works with the Wi-Fi off. One lets you
- 30:54use Cloud models without investing in
- 30:56expensive hardware. One pools multiple
- 30:58free engines together so you can switch
- 31:00when you hit a limit. And on top of all
- 31:02that, you've got a bonus daily driver
- 31:04inside VS Code that just built a working
- 31:07arcade game from a single prompt. No
- 31:09subscription, no credit card, no
- 31:11Anthropic login. Every prompt from this
- 31:14video, every base [music] block, every
- 31:16link to the tools we used, it's all
- 31:18inside my free Stay Ahead community.
- 31:20Link's in the description. Join, grab
- 31:22everything, and run your own build
- 31:24tonight. And tell us in the comments,
- 31:26which setup are you trying first?
- 31:28>> [music]
- 31:28>> Our team actually reads them, and we try
- 31:30to pick the next video based on what you
- 31:32ask for.
About this transcript
This page contains the full transcript of Stop Paying for Claude Code (3 Ways to actually use it for FREE!!) by Outskill, generated from the public captions YouTube serves with the video. The transcript has 5,668 words across 833 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.