Claude Code Desktop GRATUIT À VIE : Tuto Complet — Transcript

Full transcript

1. 0:00Well, hey everyone, it's Nathan. I hope
2. 0:01you're doing well. In this video, I'm
3. 0:04going to show you how to get Claude
4. 0:06Code completely for free, and there's
5. 0:08something you need to understand that
6. 0:10many people haven't grasped yet: Claude
7. 0:12Code is just a harness for AIs. That
8. 0:14means Claude Code is everything that
9. 0:16surrounds an AI model. And based on
10. 0:18that principle, while Anthropic's
11. 0:20models are indeed plugged into Claude
12. 0:23Code, you can absolutely connect any
13. 0:25other model to Claude Code to use it
14. 0:27entirely for free with your own model.
15. 0:30And before, it was impossible to do
16. 0:32this because models weren't powerful
17. 0:33enough to compete with Opus, Sonnet,
18. 0:35and so on. So using Claude Code with
19. 0:38another model was pointless because the
20. 0:39models were so bad that it just wasn't
21. 0:41feasible. So we kept using Claude Code
22. 0:44with the base models. But today, what's
23. 0:46happening is that the local models
24. 0:48being released that run on your average
25. 0:50person's computer are reaching the
26. 0:51level of Claude Opus. Meaning the best
27. 0:54model, not counting Claude Sonnet,
28. 0:56which works extremely well with Claude
29. 0:58Code, and in particular, a model like
30. 1:00Qwen 2.5 32B, which is as powerful or
31. 1:02even better than Opus. And that means
32. 1:06if you connect Qwen 2.5, which runs on
33. 1:08any average computer, to Claude Code,
34. 1:10well, you have a Claude Code with a
35. 1:12free Opus-level model for life, and you
36. 1:14can ask it as many questions as you
37. 1:15want. There are zero usage limits. I
38. 1:18wanted to show you this little chart
39. 1:20before we start the video. You can see
40. 1:23that today, frontier models, paid and
41. 1:25private models, are no longer the ones
42. 1:27winning the AI battle. And today, we
43. 1:29have open-source models like Qwen 2.5
44. 1:31Max, for example, that are reaching the
45. 1:33top spot. Now, this benchmark is from
46. 1:35last week. It evolves so fast that in
47. 1:37my opinion, within a week, it will
48. 1:38already be outdated. But this is what
49. 1:40it looks like currently. So we have a
50. 1:42model that is open-source, which anyone
51. 1:44can run. Well, not this one because
52. 1:46it's very, very large, but essentially
53. 1:47it's an open-source model that we could
54. 1:49potentially run if we had the necessary
55. 1:51computing power, and you can see it
56. 1:53beats Claude Opus. So today, saying
57. 1:55that private models are better than
58. 1:58open-source models is false. Today,
59. 2:01that's it; open-source models are just
60. 2:03as good or even better than private
61. 2:05models, so why not use them? So, we
62. 2:07have the famous Qwen 2.5 Max arriving
63. 2:09in first place, Claude 3.5 Sonnet that
64. 2:11everyone knows which is private, Qwen
65. 2:132.5 which is also open source arriving
66. 2:15in 3rd place, and the old Qwen 2.5 Max
67. 2:17version in 4th place. Here we have
68. 2:19Claude 3.5 again but the "new" version
69. 2:21this time, and Qwen 2.5 which also
70. 2:23lands in 6th place, which is another
71. 2:25open source model. Grok, Fable, all of
72. 2:27those are paid models. Qwen 2.5 Flash,
73. 2:29which I remind you runs without a
74. 2:32graphics card if you have the necessary
75. 2:34RAM and a good processor, you don't
76. 2:36even need a GPU to run it. So ideally a
77. 2:39Mac with unified memory, and it works
78. 2:41super well. So if you have a Mac with a
79. 2:44lot of RAM, you can run Qwen 2.5 Flash
80. 2:47which is in 9th place and outperforms
81. 2:50GPT-4o. So imagine the power if you
82. 2:52connect this model to Claude, you get a
83. 2:55supercharged Claude that's better than
84. 2:57Claude 3.5 Sonnet, or at least in the
85. 2:59same tier as the 3.5 Sonnet. Next, we
86. 3:02have GLM, also GLM 4 Flash, which is an
87. 3:04excellent model that, again, if you
88. 3:06have a lot of RAM, you can run. And
89. 3:09finally, we have the famous Qwen 2.5
90. 3:1132B, which is also excellent and runs
91. 3:13with only 16GB of VRAM. I made a video
92. 3:16about it if you want the details, and
93. 3:19it's better than GLM 4 and also Gemini
94. 3:211.5 Flash, and it's really not far off
95. 3:24from GPT-4o. We can see here in terms
96. 3:26of points, there isn't a huge gap
97. 3:28either. So we have a model that is on
98. 3:30the same level as Claude 3. Claude. So
99. 3:34how do you do it? The first thing you
100. 3:36need to do is download the Claude
101. 3:37application. You download it to your
102. 3:39computer and you end up here. We don't
103. 3:41want to pay, so we won't create an
104. 3:43account, we won't sign up, we won't
105. 3:44enter anything at all, we're just going
106. 3:45to use Claude for free. So to do that,
107. 3:48you'll need to go to the developer menu
108. 3:49. But be careful, you don't have this
109. 3:51menu by default. To be able to activate
110. 3:53it, you have to go to Help and then
111. 3:54Troubleshooting, and here in the list,
112. 3:56you will have the option "Activate
113. 3:58developer menu." I don't have it
114. 3:59because I already activated it, but you
115. 4:01will see "Activate developer option"
116. 4:03here in the list. So you just have to
117. 4:04click on it, and it will restart Claude
118. 4:06. And then you reopen it, you arrive
119. 4:07again at this window and there you have
120. 4:09the famous developer menu that appears
121. 4:11right at the top. On Windows, it's
122. 4:12pretty much the same principle. You
123. 4:13have the little burger menu that shows
124. 4:15you the same options you have here on
125. 4:16Mac. Once you have this menu appear,
126. 4:18you just need to click on configure
127. 4:20third-party inference. So you click on
128. 4:22it and here you will be able to
129. 4:23configure your external model. So, I
130. 4:26had already done a configuration, but I
131. 4:27will show you from scratch how to do it
132. 4:29. So first you need your own AI server.
133. 4:32For that, the simplest way is to use
134. 4:33the Agent application. I will put the
135. 4:34link for you in the description. For
136. 4:36all those who have been following the
137. 4:37channel for a long time, you know this
138. 4:38application very well. It is developed
139. 4:39by me and by the community. Now, I have
140. 4:41many contributors, many people who work
141. 4:43on the code and that makes me extremely
142. 4:44happy. So the project is becoming
143. 4:46really interesting and it's the
144. 4:47simplest way to create your little
145. 4:49server at home. So what I'm going to do
146. 4:51now is head over to—well, I have
147. 4:53access to my server via Agent Link. By
148. 4:55the way, for all the people who haven't
149. 4:57created an Agent Link account yet, here
150. 4:59is what it looks like. So you have help
151. 5:02, which is several pages explaining how
152. 5:03to do this or that thing. Basically,
153. 5:06it's tutorials. And you also have
154. 5:07support, where you can chat directly
155. 5:09with me, and I answer your questions if
156. 5:11you have any, and I help you deploy
157. 5:13your chat live with me. And you also
158. 5:14have the list of servers, of course. So
159. 5:16once you have connected your server,
160. 5:18you can click on it and access it
161. 5:20remotely from anywhere. So now I am on
162. 5:22Agent. What I am going to do is head
163. 5:25right here. I am going to go to API and
164. 5:27here, I have enabled public exposure.
165. 5:29So I will be able to connect to my
166. 5:31Agent server, which is not at my home,
167. 5:32from anywhere. So here, I am on my Mac
168. 5:34and my Agent server is in another
169. 5:35location, but I can still access it
170. 5:36thanks to Agent Link. Otherwise, it
171. 5:38will be the local network URL if you
172. 5:40are on the same network as where your
173. 5:42AI server is. So here, what I am going
174. 5:45to do is take the API URL of my AI
175. 5:47server and enter it right here. Now be
176. 5:49careful, for Continue, you must remove
177. 5:51the/v1 otherwise it won't work. And for
178. 5:53the identifier type, make sure to set
179. 5:55it to static API key. Next, for the API
180. 5:57key, if you have Agent, you have it
181. 5:59right here. So copy it and paste it
182. 6:01right here. For the authentication
183. 6:02schema, make sure to check Bearer. Next
184. 6:04, right here, you'll see model
185. 6:06discovery. You need to uncheck that,
186. 6:07because by default, it doesn't discover
187. 6:10the right model. So uncheck it, then go
188. 6:12here, click add, and enter your model's
189. 6:15configuration. For the model ID, I
190. 6:17recommend using local cloud so you
191. 6:19won't have any issues. For the display
192. 6:21name, we'll keep Command 3.8 27B, but
193. 6:23you can put whatever you want. If you
194. 6:25want, name it Opus 4.6 so it feels like
195. 6:26you're using Opus 4. But we'll keep
196. 6:29Command 3.8 27B here, since that's the
197. 6:31model running in the background. For
198. 6:33the tier alias, you can put Opus, since
199. 6:35it is an Opus-tier model. You can also
200. 6:37display the estimated cost. This will
201. 6:40let you see how many euros you've saved
202. 6:42, because it shows what it would cost
203. 6:44if you were paying for Claude, even
204. 6:45though it's totally free for you since
205. 6:47it's running on your own server. You'll
206. 6:50just have to pay for the electricity,
207. 6:52but that's nothing, less than 10 euros
208. 6:54a month. It all depends on your
209. 6:55hardware, your PC, etc., but generally,
210. 6:57power consumption for AI remains very
211. 6:59low. Once that's done, you can test the
212. 7:01connection right here. When the little
213. 7:03light turns green, it's perfect. You
214. 7:05can then apply the changes. So now I'm
215. 7:07going to rename it Opus 4. For the
216. 7:10example, apply changes, save, and
217. 7:12restart. And Claude will automatically
218. 7:14restart. There you go, Claude is open;
219. 7:16I didn't even log in, yet I have full
220. 7:18access to Claude. So, I have the model
221. 7:19name appearing right here. It says Opus
222. 7:214.6 because that's the name I gave it,
223. 7:23but you could have put 3.8 27B. And now
224. 7:25, I'm ready to work. But wait, just
225. 7:27before that, let's look at my system
226. 7:29prompts. A system prompt that is huge.
227. 7:32It's nearly 20,000 tokens. So if your
228. 7:34context window is too small like with
229. 7:36this model, it won't work. It works
230. 7:38very well with a 65k context on Agent,
231. 7:41but it doesn't work well with Claude.
232. 7:43In any case, you will be limited. So if
233. 7:45you're using Continue, I recommend
234. 7:47using a large context instead. For
235. 7:49example, I have this Q4K XL model here
236. 7:51that has 135k context. So we'll be much
237. 7:53more comfortable with that. I'll just
238. 7:54select it. I'll switch to this model.
239. 7:57I'll wait for it to load. There we go,
240. 7:59the model is ready and loaded. I'll
241. 8:00also show you the configuration I've
242. 8:02set up. We're using the Qwen 2.5 7B Q4K
243. 8:05XL. I put it on both cards because I
244. 8:07have two graphics cards. We have
245. 8:09135,000 for the context. For the
246. 8:10offload mode, I've set it to layer mode
247. 8:12. I'll make a dedicated video for that
248. 8:13later, but it significantly speeds up
249. 8:15the generation rate. It adds
250. 8:16practically 10 tokens per second. So I
251. 8:19recommend enabling this mode if you're
252. 8:20using two graphics cards. For the cache
253. 8:22, we're on Q5_1; it's practically the
254. 8:24same as Q8 but saves a bit of space. We
255. 8:26enable MTP, and the most important part
256. 8:29is the configuration. I'm just using a
257. 8:32custom chat template here that allows
258. 8:34it to work with Continue, because by
259. 8:36default, the Qwen 2.5 chat template
260. 8:38doesn't work with Continue since it
261. 8:39places system messages within the
262. 8:41conversation, whereas Continue requires
263. 8:43the system message to be strictly at
264. 8:45the beginning and never afterwards.
265. 8:48Because of that chat template format,
266. 8:51the base Qwen 2.5 template doesn't work
267. 8:53. That's why I created a corrected chat
268. 8:56template to work with Continue. I'll
269. 8:58provide it for download in the
270. 9:00description so you can get it. If
271. 9:02you're in the settings, you just need
272. 9:04to add this line, the template engine,
273. 9:06to enable the chat template, and for
274. 9:08the chat template file, load the one
275. 9:10you downloaded to ensure it's
276. 9:11compatible with Continue. I'll let you
277. 9:14copy my configuration if you'd like.
278. 9:15Once that's done, everything is
279. 9:17configured. We don't need the settings
280. 9:18anymore. We can go straight to Continue
281. 9:20, say hi, and see if it responds. It
282. 9:23takes a little while to start up at
283. 9:24first, about a minute; that's normal.
284. 9:26It's because it's reading the entire
285. 9:27massive system prompt from Continue. If
286. 9:29it takes time to start at the beginning
287. 9:30, it's totally normal. There, it's
288. 9:32responding. There we go. Hi, what can I
289. 9:33do for you? So let's test it out. We'll
290. 9:35tell it to create a test.txt file on
291. 9:38the desktop and tell people to
292. 9:41subscribe. There we go, let's get
293. 9:43started. It's generating that right now
294. 9:44. As you can see, it's really close.
295. 9:46Just like if I were using Claude paid,
296. 9:47even though I'm not at all. I'm using a
297. 9:49Claude 3.8 behind the scenes that I'm
298. 9:50absolutely not paying for and that runs
299. 9:52locally on my server. So there it is,
300. 9:53it's creating the test file. It's
301. 9:55exactly the same user experience as if
302. 9:56I were paying for the other Claude.
303. 9:58It's done. It created the test.txt file
304. 10:00on my desktop where it says that it
305. 10:01obviously has no way to actually
306. 10:03contact people. But that's normal, it
307. 10:04was just for a test. I'm going to tell
308. 10:06it that, let's see what it replies. It
309. 10:07was just for a test. To see if it works
310. 10:11well. Sure, I created the file without
311. 10:13any issues. Do you want me to delete it
312. 10:14or leave it? Let's verify it first
313. 10:16before it deletes it. So indeed, it
314. 10:18created it right here. Wait, it didn't
315. 10:19put the text inside. That's normal
316. 10:21because it thought I really wanted to
317. 10:22send that to people. Go write a message
318. 10:26in this file telling people to
319. 10:29subscribe. It's written don't forget to
320. 10:32subscribe. So yes, it correctly put "
321. 10:33don't forget to subscribe" directly in
322. 10:35the file. So it's working correctly. I
323. 10:36can even tell it to create a landing
324. 10:40page to present Claude in an HTML file.
325. 10:44So here, we even have access to
326. 10:46Anthropic's skills. So it will use its
327. 10:48frontend design skill to make a nice
328. 10:50design. We really have code that works
329. 10:52perfectly. There we go, the generation
330. 10:53is complete. So, it wrote about 400
331. 10:55lines of code and coded this little
332. 10:57landing page that briefly presents the
333. 10:59Claude Opus model. The site is
334. 11:01perfectly responsive, and it's quite
335. 11:03nice. Well, nothing too special, it's
336. 11:05just a very basic little presentation
337. 11:07of Claude Opus. I didn't ask for
338. 11:09anything more. And we have a site that
339. 11:11is beautiful, functional, and does
340. 11:12exactly what I asked. So that's it for
341. 11:14the example. You can now use Claude
342. 11:16code entirely for free thanks to this
343. 11:19technique. And obviously, when there
344. 11:21are new, even more powerful models, or
345. 11:23even if you want to connect other
346. 11:24models, you can do it directly from the
347. 11:26agent. You just have to create presets,
348. 11:28and when you change presets, you don't
349. 11:30need to reconfigure everything. You
350. 11:32have the model that updates
351. 11:33automatically on Continue. For example,
352. 11:35if I switch here to the Q27BQ6K MTP, so
353. 11:37the slightly less quantized model,
354. 11:39which is a bit smarter, it will switch
355. 11:41over. I could keep talking here
356. 11:43directly in Continue. I don't need to
357. 11:45reconfigure it every time. That's why
358. 11:46it's best to use your provider's name
359. 11:48for the model name. So if you're on
360. 11:50Ollama, write Ollama. That way, you
361. 11:51know it's connected directly to Ollama,
362. 11:53and you can change the models later
363. 11:54right inside Ollama. There, I'm done
364. 11:56with this short tutorial. I'm not going
365. 11:57into details here because I've already
366. 11:59covered the model and Ollama
367. 12:01configuration in other videos. I'll
368. 12:03make a video dedicated to Ollama soon
369. 12:04so you can see how to install it,
370. 12:06configure it, and so on, but there's
371. 12:07nothing very complicated. You go to the
372. 12:09site at the bottom, download the .exe
373. 12:11directly, click on it, it opens on
374. 12:12Windows, and you just have to do your
375. 12:14configurations, following the videos
376. 12:15I've already made on the subject to set
377. 12:17it up. So there you go, I'll let you
378. 12:19have fun with Continue. Don't hesitate
379. 12:21to give me your feedback on your best
380. 12:22settings and how you use it. Does it
381. 12:24work well? And also, for those who have
382. 12:26a slightly less powerful PC, I
383. 12:27recommend using Ollama directly without
384. 12:29going through Continue because
385. 12:31everything is integrated into Ollama.
386. 12:33You have the same functionality as
387. 12:35Continue. You have projects, memory,
388. 12:37internet access; you have everything
389. 12:39included. So just talk directly with
390. 12:41Ollama right here. It also works
391. 12:42extremely well. You don't necessarily
392. 12:44need to connect it to Continue because
393. 12:46Continue requires having a large
394. 12:47context. So, I hope you enjoyed this
395. 12:49short video. Don't forget to like,
396. 12:50subscribe, and share if you liked the
397. 12:52video. And we'll see you very soon in
398. 12:54the next video. This is Nathan Nine Tus
399. 12:56.