Alternatives gratuïtes a Wispr Flow, ElevenLabs i CapCut: 13 projectes de GitHub
Una selecció de tretze projectes oberts per editar vídeo amb agents, dictar i clonar veu en local, donar memòria al codi, dissenyar, vigilar esdeveniments globals i automatitzar recerca.
Tretze projectes oberts que van més enllà de la moda
Andrew Warner i Adam Brakhane repassen en aquest episodi una selecció de repositoris que han guanyat visibilitat a GitHub. El fil conductor no és només quantes estrelles acumulen, sinó què permeten construir: vídeo controlat per agents, dictat local, veu sintètica, memòria estructural per a bases de codi, navegadors amb agents i alternatives obertes a plataformes de disseny o edició.
La llista barreja eines madures amb prototips recents. Per això convé entendre-la com una curació editorial, no com un rànquing científic ni una recomanació d’instal·lació automàtica. Alguns repositoris processen dades sensibles, demanen claus d’API o controlen comptes externs. Abans d’executar-los, cal revisar la llicència, els permisos, les incidències obertes i l’activitat real del projecte.
1. OpenMontage: producció de vídeo dirigida per un agent
OpenMontage vol convertir un assistent de programació en un estudi de producció. L’usuari descriu el vídeo en llenguatge natural i l’agent pot investigar, preparar el guió, buscar o generar recursos, muntar-los i renderitzar el resultat. Inclou fluxos per crear explicadors, peces amb imatges generades i muntatges a partir de vídeo d’arxiu.
El debat del vídeo destaca una diferència important respecte d’un editor tradicional: l’agent pot omplir els espais entre eines i automatitzar una cadena completa. El repositori funciona amb assistents capaços de llegir fitxers i executar Python, entre els quals Claude Code, Cursor, Copilot, Windsurf i Codex. Necessita també dependències com FFmpeg i Node.js, i alguns fluxos poden generar costos si utilitzen serveis externs d’imatge, veu o vídeo.
2. codebase-memory-mcp: un mapa persistent del codi
codebase-memory-mcp indexa un repositori i en construeix un graf de coneixement amb funcions, classes, rutes i relacions. Després exposa consultes estructurals mitjançant MCP perquè un agent pugui preguntar què crida una funció, quin impacte tindria un canvi o com es connecten diversos serveis sense rellegir milers de fitxers a cada torn.
El projecte afirma que redueix dràsticament els tokens necessaris per explorar codi i que funciona localment. Són xifres dels seus propis benchmarks, no una garantia universal: el guany dependrà del llenguatge, la mida del repositori i la mena de pregunta. També és una eina amb accés ampli al codi i que pot modificar configuracions d’agents, de manera que és prudent auditar-la abans d’instal·lar-la en projectes privats.
3. TimesFM: prediccions per a sèries temporals
TimesFM, de Google Research, és un model fundacional entrenat per pronosticar sèries temporals. Rep una seqüència numèrica —vendes, demanda, sensors o altres mesures ordenades en el temps— i calcula valors futurs. El repositori obert permet executar el model, ajustar-lo i integrar-lo en fluxos de Python.
És una eina especialitzada, no un oracle general. La qualitat depèn de la distribució de les dades, dels canvis de règim i de l’horitzó de predicció. El mateix repositori aclareix que l’edició oberta no és un producte oficial amb suport de Google. Pot servir com a punt de partida tècnic, però qualsevol decisió important necessita validació amb dades pròpies i comparació amb baselines més simples.
4. peerd: un agent que viu dins del navegador
peerd és una extensió que executa el bucle de l’agent al mateix navegador. Pot operar pestanyes, utilitzar càlcul local i treballar amb el model triat per l’usuari, sense un backend propi ni telemetria declarada. La proposta s’adreça a qui vol una experiència semblant als assistents integrats al navegador però amb més control sobre el proveïdor i les dades.
L’execució local redueix dependències externes, però una extensió amb capacitat per llegir i manipular pestanyes continua sent molt poderosa. Cal revisar quins permisos demana, separar perfils sensibles i entendre on s’envien les peticions del model configurat. «Sense backend» no vol dir necessàriament «sense connexions»: si es tria un model al núvol, el contingut necessari arribarà al seu proveïdor.
5. FluidVoice: dictat local per a macOS
FluidVoice és una aplicació de dictat fora de línia per a macOS. Transcriu la veu localment i insereix el text a altres aplicacions, amb la intenció de ser una alternativa gratuïta a serveis com Wispr Flow o Superwhisper. Un dels presentadors la prova després de gravar l’episodi i explica que, superat un dubte inicial, li funciona prou bé per plantejar-se cancel·lar una subscripció.
Aquest entusiasme és una impressió puntual, no una comparativa controlada. La precisió, la latència i el consum de recursos depenen del Mac, del micròfon, de l’accent i del soroll. El benefici clar és que l’àudio i la transcripció poden quedar a la màquina, sense una quota mensual ni l’enviament constant de veu al núvol.
6. birdclaw: una memòria local per a X
birdclaw, de Peter Steinberger, importa arxius de X/Twitter i organitza publicacions, missatges directes, m’agrades i marcadors en una base SQLite local. Ofereix una interfície més enfocada, cerca completa i sortides estructurades perquè agents i automatitzacions puguin treballar amb les dades.
El projecte es presenta explícitament com a treball en curs. Algunes lectures i escriptures en directe depenen d’eines addicionals o de sessions del navegador, i aquests mecanismes poden trencar-se quan X canvia la plataforma. La idea de recuperar un arxiu propi i fer-lo consultable és sòlida; la interacció automatitzada amb el servei requereix més prudència pels permisos, les credencials i les condicions d’ús.
7. World Monitor: un quadre global d’informació oberta
World Monitor agrega notícies, geopolítica, activitat militar, mercats, ciberamenaces, clima, vaixells i aviació en un mapa i diversos panells. L’objectiu és acostar a qualsevol usuari una visió integrada que recorda eines d’intel·ligència molt més costoses. El projecte ofereix variants centrades en món, tecnologia i finances, i es pot executar com a web o aplicació d’escriptori.
La presentació impressiona per la densitat de capes i fonts, però un tauler no converteix dades heterogènies en veritat automàticament. Les alertes poden arribar tard, duplicades o sense context; les síntesis generades per IA també poden equivocar-se. És útil per descobrir senyals i orientar una investigació, no per prendre decisions operatives sense comprovar la font original.
8. Penpot: disseny obert i preparat per al codi
Penpot és l’alternativa més consolidada de la llista. És una plataforma oberta de disseny i prototipatge que utilitza estàndards web i es pot autoallotjar. El vídeo la contraposa a Figma per evitar que els fitxers i el flux de treball quedin lligats a una única subscripció.
A més del repositori principal, Penpot manté un servidor MCP oficial que permet consultar i modificar un disseny des d’un client compatible. Això obre fluxos de disseny a codi i de codi a disseny, tot i que no elimina la feina de sistema visual, accessibilitat o revisió humana. El principal atractiu no és només que sigui gratuït, sinó que el format i la infraestructura poden quedar sota el control de l’equip.
9. Voicebox: clonació, síntesi i dictat en local
Voicebox reuneix les dues meitats del flux de veu: pot clonar una veu, generar àudio a partir de text i dictar en qualsevol camp. Funciona localment, integra diversos motors de síntesi, ofereix una API i un servidor MCP i es presenta com una alternativa conjunta a ElevenLabs i Wispr Flow.
La versatilitat té un cost tècnic. Alguns models ocupen gigabytes, la primera arrencada és lenta i el rendiment varia molt entre CPU, Apple Silicon i GPU. La clonació de veu també exigeix consentiment: que una eina sigui local i oberta no legitima imitar terceres persones. Per a veus pròpies, accessibilitat o producció controlada, en canvi, pot evitar quotes i conservar els enregistraments a casa.
10. system_prompts_leaks: estudiar com s’instrueixen els agents
system_prompts_leaks recopila prompts de sistema i configuracions atribuïdes a assistents, editors i agents comercials. Els presentadors remarquen que «filtrat» és una etiqueta massa dramàtica: part del material prové de documentació pública o ha estat revelat pels mateixos fabricants, mentre que altres fitxers s’han extret i poden ser incomplets.
La col·lecció és útil per estudiar estructures, criteris d’eines, límits i tècniques de prompting. No s’ha de tractar com una font immutable ni copiar-la a cegues. Els productes canvien sovint, l’autenticitat pot variar i un prompt sense el model, les eines i la infraestructura originals no reprodueix el mateix comportament.
11. Agent-Reach: donar accés web als agents
Agent-Reach automatitza la instal·lació i comprovació d’eines perquè un agent pugui llegir o cercar a X, Reddit, YouTube, GitHub, LinkedIn, RSS i altres fonts. No crea una nova capa universal: configura utilitats existents com yt-dlp, el client de GitHub o lectors específics, i després l’agent les invoca directament.
Aquesta arquitectura és senzilla i intercanviable, però no fa desaparèixer els bloquejos, les cookies ni els canvis d’API. Cada canal conserva els seus requisits i riscos. És especialment recomanable executar el diagnòstic, activar només les fonts necessàries i no lliurar una sessió personal completa a un agent si una credencial més limitada és suficient.
12. daily_stock_analysis: informes financers automàtics
daily_stock_analysis analitza una llista de valors, combina dades de mercat i notícies i genera un tauler amb senyals com comprar, mantenir o vendre. Pot programar informes, enviar notificacions i desplegar-se amb Docker. El projecte també incorpora proves retrospectatives per comparar recomanacions passades amb moviments posteriors.
El vídeo l’identifica com una tendència creixent: eines personals que substitueixen fulls de càlcul per fluxos d’IA. Això no converteix el resultat en assessorament financer. Les dades poden ser incompletes, els models poden inventar una explicació plausible i un backtest pot patir biaixos. És millor veure’l com una plataforma experimental de recerca que com un pilot automàtic per invertir diners.
13. OpenCut: l’alternativa a CapCut encara s’està reconstruint
OpenCut aspira a ser un editor gratuït i obert per a web, escriptori i mòbil. El pla actual inclou un nucli en Rust, API d’edició, plugins, mode sense interfície i un servidor MCP perquè els agents puguin controlar el muntatge. La idea encaixa amb la tendència de l’episodi: no només editar manualment, sinó oferir infraestructura programable.
Aquí els presentadors són més prudents. L’edició de vídeo és difícil, els fitxers són grans i la previsualització, els efectes i l’exportació requereixen molta enginyeria. El projecte s’està reescrivint i el seu propi README deriva els usuaris cap a la versió clàssica mentre la nova no està preparada. OpenCut té activitat i suport, però encara li falta camí per substituir CapCut en un flux exigent.
L’aparador patrocinat: Zapier MCP
Al mig de la selecció hi ha un segment patrocinat de Zapier MCP. Andrew explica que prefereix aquesta capa per connectar agents amb correu, calendari i altres aplicacions en lloc d’exposar cada servei directament. Zapier ofereix permisos per aplicació i acció, historial i connexions gestionades.
És una opció còmoda, no una peça de codi obert de la llista. Les crides consumeixen quota del pla de Zapier i continuen requerint una configuració de mínim privilegi. L’avantatge és centralitzar credencials i controls; el desavantatge és introduir un intermediari comercial en fluxos que alguns dels altres projectes intenten mantenir locals.
Com valorar un repositori abans d’adoptar-lo
La part final aporta un criteri útil. Quan una eina dona senyals dolents —una demo caiguda o comportament inestable— Adam mira qui la finança i quan s’han actualitzat els fitxers. Un patrocinador fiable ajuda, però diversos suports independents redueixen la dependència d’un únic actor. L’activitat recent tampoc garanteix qualitat, però un repositori abandonat és un risc si el producte serà central per al negoci.
Andrew explica que dedica hores a preparar l’informe encara que Claude Code l’ajudi a reunir informació. Revisa si algú intenta manipular la selecció, comprova materials i cura les demostracions. Aquesta és també la millor manera d’utilitzar la llista: deixar que l’agent acceleri la recerca, però conservar el judici humà sobre seguretat, manteniment i utilitat real.
Les estrelles indiquen atenció, no estabilitat. Abans d’incorporar qualsevol d’aquests projectes convé provar-lo amb dades no sensibles, fixar una versió, revisar la llicència, comprovar releases i incidències i preparar una sortida. El valor de l’ecosistema obert és poder inspeccionar i adaptar el programari; aprofitar-lo bé exigeix precisament fer aquesta inspecció.
Contrast i context
Fonts consultades
-
01
The Next New Thing Recursos de GitHub de l'episodi
-
02
GitHub calesthio/OpenMontage
-
03
GitHub DeusData/codebase-memory-mcp
-
04
GitHub google-research/timesfm
-
05
GitHub NotASithLord/peerd
-
06
GitHub altic-dev/FluidVoice
-
07
GitHub steipete/birdclaw
-
08
GitHub koala73/worldmonitor
-
09
GitHub penpot/penpot
-
10
GitHub jamiepine/voicebox
- 11
-
12
GitHub Panniantong/Agent-Reach
- 13
-
14
GitHub OpenCut-app/OpenCut
-
15
Zapier Zapier MCP
Font de treball
Transcripció amb marques de temps
Consulta la transcripció
-
0:00
, obre el vídeo en una pestanya nova
I got a WhisperFlow alternative that's free, self-hosted, and super fast coming up. I've got an AI tool that with one prompt will create a beautiful video for you with no AI slop, I guarantee it. And I'm going to show you a tool that will clone your voice so free your mom won't even know it's not real. All that and so much more coming up. >> [music] >> Presented by Zapier, the AI automation company. >> All right, Adam. People freaking love this thing. They're actually using the repos or they're learning about how to build with them. And so, let's teach them what the top 10 trending GitHub repos are for this week. Number one, and I like to start with number one, not make people wait till the end of the list to get to it.
-
0:27
, obre el vídeo en una pestanya nova
Number one is Open Montage. Here it is. You describe a video in plain language and your coding agent will research it, will script it, will edit it, will render the whole thing. Usually, I like to click directly into the GitHub and show you around so that I can get your feedback. actually want to show you what this looks like first and then let's get your feedback. Here is a YouTube video showing what it looks like in the end. Look at this. Really nice, right? I'll show you the prompt in a moment. I'm not going to play the whole video cuz this thing just goes on and on and on. But, it's like nice, it's dramatic, it doesn't look like it's AI generated, it doesn't feel cheesy. Look at this. It merges well together. Here's the prompt that he used to create it. >> is actually without any of the video API keys. So, which is awesome. So, um once you run clock code and put your prompt inside the cloud, and then you can see that after uh running the prompt, for example, we're making a 75-second documentary montage about sea life in the rain at night. So, once I've >> That's it. It's like two sentences that he has in there. Are you able to hear him talk by the way, Adam? >> Uh yeah, I am, yeah. >> Excellent. Two sentences, he gets the whole thing. It's done on his computer. He specifically wants no AI generated images, he wants real footage, and he got it. What do you think about this one? >> I mean, that explains why it doesn't look AI generated because it's it's all it's like stock footage, right? Like it's not AI-generated. It's just cutting that together. I mean, I think that's interesting for a dramatic clip of rain falling with cool music, but but I also think it's it's like why would I use this? Andrew, you've used a lot of these tools. Why would I use this instead of Descript or CapCut or something like that? Where What gap does this fill? >> Uh CapCut does have some of these features in it. Descript is having some of these features in it. Neither one of them is as fully baked as this. And then also, Adam, what I'm seeing is a lot of people are using Claude code to create videos, whether it's it's edit their videos of themselves. And a lot of them are using like text over their screen or images, but they're not using B-roll. Or they just want just standard B-roll. And um and even 11 Labs' voices. And what this does >> Andrew, I think I'm I'm frustrated in like Descript and CapCut. I'm usually frustrated by the experience. You know, they built these nice interfaces, they built AI into them, but they're still very limited. And when I use Claude code, I don't feel limited by anything. So, I actually have started to prefer tools like this that plug into Claude code and let me combine tools and use part of this tool, part of that tool. >> Right. >> And now I'm not limited to this like one editor experience that oh, it doesn't have that one little feature I needed to finish the video. Claude code is like oh, write me a little script that can build that tool that I want really quickly. And it says great, I'll I'll do that. So, I I love these, cuz they supercharge kind of the Claude code experience by not requiring that you have to rebuild this stuff from scratch. They've done a lot of the work. Uh but then it's still flexible in a way where I could build anything that I could possibly dream of. >> More and more of my friends are using Claude code to edit their videos. Um I like it. I think this is really interesting. No wonder it's number one. It works well. I should also say that he specifically wanted it all on his desktop, and he didn't want to pay for API keys. But you can, if you decide that you do want to have external tools, if you do want to have AI generated images, this thing does it. It's very robust. It's a really good uh experience. And no wonder it's number one for this week. Let's go on. What's number two?
-
4:03
, obre el vídeo en una pestanya nova
Number two is Code Based Memory MCP. Boy, this report is getting longer and longer because I also wanted all these links in here. I wanted the Reddit cut um we'll have a link to the report below. I obviously can't anymore read everything that uh live. Um but what do you think of this? This this person by the way from on Reddit uh this is Andrew OOO. Uh Code Based Memory MCP review 99% token cut for code agents. Been running coding agents on a few repos lately and kept watching them burn through tokens just to answer questions like who calls this function or what's the impact of changing X? Got annoying fast. Found this tool. That's what he's using it for. What do you think? >> Yeah, I mean these are awesome. This is like the AI version of the the LSP tools coders have used for a really long time. Which is like how do you give your coding environment access [snorts] and understanding of even a very large uh code base. There's a lot of files, different places, and and AI is very happy to just use a hammer and change something and not realize it broke something else. It's a tool like this or an LSP that that lets the agent know, "Ooh, I need to look at this. I need to consider these things." And honestly, the the idea of coupling memory with it is something that that to me is very exciting because I struggle. If I'm working on five projects at a time, I even struggle to remember, "Ooh, I need to remind this agent to look at this file. And where was that function? And how do I want to change this feature?" I I forget. And man, I'm I'm exhausted a lot of the the time now. Like I'm a I'm a dad. Like I I've got kids. Like And I these things are are really powerful when they couple these kind of well-built tools like an LSP with memory. And then they give they give access to the those things, you know, through MCP, which is the agent's favorite way. >> For my research, I I've been adding these so what to just remind me to tell the audience why it matters. Here's Here's the so what of it. It makes an AI coding agent smarter and cheaper on big code bases. It stops >> Yeah, cheaper is a big is a big part of that. >> Um all right. Adam, as you know, we've been getting a bigger and more passionate audience. It's time for us to activate them. I need your help. This man is wanted, Matt Van Horn. He is a creator we've talked about before. He's got two projects here on GitHub. One of them actually trended and made it into our list. One did not, but is still good. And I want to have him on here to talk about them. He is ignoring me, but he's a nice guy, and I'm sure that he's getting a lot of people who are reaching out to him, so I'm not blaming him. But if anyone out there knows him and wants to support the show here, reach out to him. And if you don't know him, reach out to him on X. And if you don't know him and don't want to reach out to him on X, maybe you can throw me a like and a subscribe. I'm really trying to build this up. And let me let me also, just to show you what a nice guy he is beyond building this. This is an experience, Adam, I think you could relate to. Was this from this morning? >> Yeah. No, from yesterday. He goes, "I love open source. Freddy Montero, soccer star player from Colombia and Seattle Sounders number two all-time goal scorer, submitted two PRs, two pull requests, both merged into / 30 days." Now, listen to this. Listen to this. This is where the story I think is so relatable to you. He texted me last week, quote, "Can I get breakfast and stop by your office? You helped me out to set up Cloud Code." Matt did, and now he's shipping PRs. Isn't this like your experience where you got me excited and then I started building and now I'm coming back in and showing you what I did? >> [snorts] >> That's amazing. >> All right, guys, if you're out there, help me get Matt. And if you're not, >> Matt, we know you're out there. We want to talk to you. >> Yeah. >> Maybe we should submit some some PRs to to to get on his radar. >> That might be the next step. Okay. Boy, this really is going on too long on the page. All right, um
-
7:39
, obre el vídeo en una pestanya nova
This one is from Google research. It's times F M A It You feed it a sequence of numbers like what your sales data is, what the demand is, what the traffic how your ads are performing, anything. And this is what got controversy on Hacker News and I'll show it in a moment. Then it predicts what comes next. It does not need to be trained to do this. I'm going to show you the GitHub repo right here, then I'll show you the controversy on Hacker News. People were shocked shocked that this actually can work and didn't believe it. Let's take a look at it. Um I somehow find the concept of a general time series model strange. How can the same model predict egg prices in Italy and global inflation in in a reliable way? And how would you even use this model given that there are no explanations that help you trust where the prediction comes from? And people who came in there and started explaining it. What What is not generally understood is that these models don't predict egg prices or inflation in Italy. They decompose a time series of trends, seasonality, and residuals. That's why they're act That's what That's what they're actually modeling. They cannot predict wars in the Middle East influencing inflation unless there is a seasonal pattern. And then people say, "Well, actually seasonal pattern for wars in the Middle East, maybe." Here's a Here's a real use case that will show everybody how you might be able to incorporate this by a Googler. Uh he says, "I when I worked on Google Ads, we used time series forecasting to compute the odds of an ad campaign reaching its goal to tell users how likely they were to hit their goals. See, a ton of unsophisticated advertisers would just draw line from zero to the number they were at today, project then they would project that line to the end of the month, and they would consider that the forecast of the amount of conversions and spend that they were going to hit. This, of course, doesn't take into account various seasonalities, the time of day, time of year, etc. And it gives them a poor forecast. That's what this is for. Is it perfectly accurate? >> I have a I have a confession to make. >> Yes, sir. >> I've been guilty of this. We've been guilty of this. [laughter] You know, it's the only thing you can do reasonably in a spreadsheet. You look at the number and you know it's not right, but you're like, "Ah, it's about the best I can do in in Google Sheets if that's where we're, you know, reporting." So, sorry, guilty. That's that's me. I'm I'm I'm very excited to to see this coming from them though, cuz we've got all these language models and I thought for a long time now, you know, old AI AI been around for for decades. We've been using it for decades. Facebook is AI, all ads are AI, uh you know, the model underneath, but it's machine learning and it's numerical AI. And this new AI is all text AI. And text AI is not very good at math. It's good at storytelling. It's good at understanding concepts. And so, I've been waiting for these two things to combine with each other to see, well, how can I feed whatever data I want without having to learn how to train my own machine learning model? Cuz today, if you want to do machine learning and you want to do your own forecasting, you're going to need a data scientist. You're going to need a data engineer. That's still just as complicated as it was 10 years ago. So, I'm really excited to see this stuff coming out of Google and I'm hoping that this is the first of like many reports, many uh research articles that they put out and that this will be a wave of numerical AI kind of for the masses, for you and me. >> I love this. I love that we could incorporate it. I love that we have access to it. Thank you, Google. All right, that's the third most popular or third trending repo of the week according to GitHub. I want to do a quick mention here of Zapier MCP. Adam, I got to I got to tell you a quick story. This week on X, there was this trending story by a Googler who goes, "I was fired for giving people access to Google Workspace using CLI." Because what he wanted to do was say, "All these people like Andrew are using Hermes Agent, Open Claw, etc. I want to give them access so that their agents get access to all these tools." I messaged him and now that we're getting a little bit of clout, plus he's a nice guy, I'm a nice guy, I say, "Could you please come on and do do a video with me showing me how this works?" He comes on with me and goes, "Andrew, I want to show you, but my setup is like in between. I'm now re- re-doing my whole system here. I can show you, but not that much." I said, "How about if I show if I connect my Hermes agent to all these tools?" Cuz I want to give my Hermes agent email, calendar, and so on. He goes, "Yeah." >> Yeah. >> Dude, the guy who built it gets on on a call with me. We spend over an hour trying to connect it with Google Cloud. And everyone who I talked to about Google Cloud goes, "Ah, I hate that thing." We had to do that. We had to do so many things. By the end of over an hour working with the guy who created this, who everyone was heralding for for adding more access to agents so that they could use Gmail, calendar, Asana, all that. I mean, Google uh Gmail, calendar, etc. I couldn't make it work. >> [laughter] >> I go, "Oh, but now I really want this." So, you know what I did? I go to zapier.mcp. zapier.com/mcp. No exaggeration. I connect it in and I say, "Not only give it Gmail, Google Calendar, but also uh I give it access to I don't use Asana, to Notion, etc." It happened in less than freaking 5 minutes. Now, Adam, you might say to me, "Couldn't you just tell Hermes to connect to it directly?" Yes and no. First of all, it's not so simple to connect to it directly, number one. I don't want to use the browser, number two. And number three, freaking Hermes agent goes stupid sometimes. And I don't want it to go stupid on apps that matter to me and my team. That's why I use zapier.mcp. If you're out there and you haven't tried it, zapier.com/mpcp. I went long with this cuz it's coming from a place of passion. If you're building, give your tools access to all your other tools that you have. And do it in a reliable way that won't just take for freaking ever. All right, let's go on. Since we got Zapier, we are now going to give three special mentions to uh repos that are not trending, but I'm noticing people talking to them. We're going to break into our top 10 list with these three.
-
13:21
, obre el vídeo en una pestanya nova
This is Peered. It's an AI agent that lives in your browser. No back end, no telemetry, no going anywhere. Here's what it does. Many of us have used Gemini in the Chrome browser, where you get on the right side, you start asking questions. What if you don't want to use Gemini? What if you want to use a different model? What if you want to use your own self-hosted model? That's what this guy created. It was on Hacker News, and I thought we'd show it. What do you think of this one? >> [snorts] >> I think it's awesome. I think this is This is one more of those bridges that I'm excited about. Local AI models, today they're okay, but they're they're you know, 8 months behind. The best open source models are about 8 months behind the frontier. And today's frontier models are great. So, in 8 months >> I don't know if that's true. >> It's I don't know if it's true. >> Really? >> GLM, the Oh, yeah, have you tried the 5.2 with >> You I've been seeing benchmarks that say that it's very close to Opus 48, that it's only beaten by Fable. I've been enjoying >> they're farther off in in others. I mean, they they kind of they choose the thing they want to compete on. But let's let's say they're at most 8 months behind, right? That's that's my point is that they're very close to the performance of it. And as those are models are that you're capable of running on your, you know, laptop, or maybe you have a Mac Mini or something. Well, shoot, now I can start to run tools like this. The data stays locally on my computer or in my network. It's a lot faster cuz it's all local, and I get the control of what model do I want when, and Open AI doesn't get my data, Anthropic doesn't get my data, Google doesn't get my data. I'm pretty happy about those things. I'm I'm one of the many like DuckDuckGo users, because years ago I was like, "You know, I don't love going to google.com all the time, just telling them about my life." And so, I switched to DuckDuckGo. This this thing is like right up my alley, too. >> I switched to GLM 5.2. I love it, haven't looked back, inexpensive, it's fantastic. Folks in the comments, tell me whether Adam is right that they are 8 months behind, or whether I am right that they are neck and neck. I'd love to hear from you all. Really, some of these people are freaking smart. I mean, if they're watching this, of course they are. Okay, there we go.
-
15:27
, obre el vídeo en una pestanya nova
Next one is Fluid Voice. This is because a lot of people in the comments said to me that at one point I must have mentioned that I use something other than Whisper Flow. Well, people have been asking about it. This is like the Whisper Flow for free. All you have to do is go to GitHub and get it. Or, even if you're not a GitHub person, you can just go to their website like a normal human being, download it, and start using it. Um and I have another uh version of this later on that we'll talk about. But, here's here's from uh Dilip. He goes, "I've dictated almost everything for 6 months to Whisper Flow, 44,000 words, etc. etc. Top 1.1% of users on Whisper Flow. Last week I tried Fluid Voice, open source, runs local on my Mac, corrects as I speak with no API keys, handles slang better than I expected, canceled my paid plan. The benefit of having it on your on your computer is not just that if you say any random stuff that it won't go into the cloud, which is also a nice benefit. It's also I find faster when it's on your computer than if it has to go to the cloud, doesn't use any bandwidth. Good find. >> Oh my god, I love it. I I'm a Super Whisper pain subscriber, and today when we get off this call, I'm going to go try this, and I might cancel my Super Whisper subscription, too. I'm hoping that I can cancel it. >> Whisper Flow, 15 bucks a month. I find it to be stupid. It's stupid. And with all the competition, it's because people Anyway, folks, try it out. Let me know what you think of it. Let's go on to Time out. We recorded about an hour ago, and Adam tried Fluid Voice. Check out his response. He at first thought he had an issue, but he goes, "Never mind, it's working. It's great. I'm canceling my other subscription." All right, so there's another endorsement for Fluid Voice, and I'm glad that we get to do these sessions so he can try it out. Back to the program.
-
17:15
, obre el vídeo en una pestanya nova
One more honorable mention. This is Bird Claw from Peter Steinberger. creator of open claw. Here is the project website. You can see here how easy it is to set up. I want you to um I should also click over to GitHub. Does it help that I'm actually showing the GitHub? Uh not for this one, the repo. Um I will show you this tweet from him. I find this to actually show the clarity and the benefit of it. What we're talking about is an easier way to interact with Twitter that organizes here the screenshot that organizes the things that matter to you in the order that you want and makes it a lot let's say a lot less distracting and a lot less uh Twitter-centric and more you-centric. >> Mhm. >> Cool. >> That is cool. It it pulls the the topics and things that I'm interested in and like presents me that information, but just not without me having to go to x.com. Is that what you're telling me? [laughter] >> Uh without you having to use x.com like a like a regular human being, yes. >> That's kind of cool. >> I think it is kind of cool. People have been talking about how great this is in my little nerd circle. I thought I'd share it with the broader circle. Now, back to our 10 top 10 list.
-
18:18
, obre el vídeo en una pestanya nova
Number four, World Monitor. Listen, there are people who are running big companies who are paying for data from Palantir, who have all these dashboards, who are trying to figure out what's going on in world events and how they're impacting them, or maybe they're just trading stocks, or maybe they're just trying to make money on Polymarket. Well, you don't have to pay for all this. There is a repo on GitHub called World Monitor that puts it all right on your screen. This is the kind of dashboard that you get with wars that are going on or military activities. And I will actually, instead of showing you this, I [snorts] find the tweet to be really nice. I'm going to show you this kind of creepy video cuz I just freaking love it. I'll just play little bits of it. >> [snorts] >> Hack, cyber, and code. [laughter] They go directly into Ubuntu. Start from scratch, go into GitHub, install it. This of course will be in the report for those of you who are getting this report. Um and then you can see here military bases, nuclear sites. You can check off the box on the left to see it. >> know that I want the nuclear site data on my computer. Can I like [snorts] definitely uncheck that box so that I'm not involved in whatever weird stuff is going on there. >> Right. Um from what I understand, this is so that you could enter the prediction markets which are now getting so popular and decide what you're going to to put money on. >> You know, Andrew, something something we've started to talk about like we've got this, you know, venture studio and we build businesses. Something we've started to mention more lately is as we're evaluating an idea for a business, we'll say, "Oh, I wonder what the people on Polymarket or Kalshi think of this." You know, there's some trend that we think is going to happen and like it's a guess and we can look at news articles. We're like, "Wait a minute. Other people are already betting on this trend and we can use that and we are using it now as a great source of information for Well, what does everyone else think is going to happen? And actually like as I'm looking at this, like well, man, [snorts] that maybe I could put all of our like thesis ideas in here and have that help scrape news, but also Kalshi, Polymarket, and others to get these to get these like analytical signals. Uh you know, I was a little bit distracted by all the military base stuff, but but if I could if I could not run the military base stuff and do this other that's pretty interesting. >> No, that Look at this. Like this is just a screenshot. It's it's a little [clears throat] bit old, obviously, but it's like Israeli forces escalated strikes and clash with Hezbollah in eastern Lebanon resulting in dozens of deaths. This intensified This intensified conflict marks a significant regional destabilization. You've got Iran theater. Uh you've got live webcams from just anyone who happens to have it on in Tel Aviv in the in different parts of the Middle East. This is like just looking at one little event here. They've accumulated everything here together. Um I shouldn't say one little event in the with with everything that's going on in the world. This is one aspect of it, but man, okay, let's move on to the next one. Obviously, if you're using that you then you don't want to tell me, I understand.
-
21:18
, obre el vídeo en una pestanya nova
Let's move on to Penpot. >> [snorts] >> This is like I love this line that I came up with or at least I loved it last night at 11:00 as I was writing it. Stop paying money to rent your designs and then figure out how to get them coded. This is me trying to establish why this is the Figma alternative that is self-hosted. And the reason for that is I saw a lot of people who were using this complain that with Figma you have to keep paying them. You don't really own your designs and then getting them coded up is tough versus Penpot. Which is open source, it's an alternative, and it makes it much easier to turn what you've got into code. In fact, where's the where's the line actually that I got from them? Free self-hosted Figma alternative and every design comes out as real code. And this is done so well that they've actually raised $20 million um partially because Oh, look. Led by partially Figma's ex-COO Eric Whitman and other angels and partially it's because when Figma just blew up people said, well, what's the what's another alternative? This was taking off. >> [snorts] >> I mean, this is awesome. I I love you know, I you and I have both tried the Claude Claude design tool that they put out a a month or or so ago now and we were both I think a little bit underwhelmed by the depth and capability and you know, I don't know about you, but I ran out of credit in about 10 [laughter] minutes of of his like, oh, well, I guess I'll try again next week. And so being able to run this stuff locally, have it connect to my AI locally, maybe use a local model to do it, but honestly, I'm probably still going to connect it to GPT-55 or something. That's pretty awesome and I'm a I'm an avid Figma user, have been for a long time. It's really powerful to to prototype, but yeah, kind of feels like a walled garden now. So, this is a this is a pretty exciting thing to to try. I don't understand the name. Do you know where the name Penpot >> Penpot, [snorts] no. Um, this is less LLM building and more human beings building. >> Yeah, but it connects. I saw in there there's an MCP server, so I can build in here. It's like replacing the the UI of Figma, which I love, but then I can like connect my agent to it through MCP to presumably vacuum out all of that information, help code, and and stuff like that, right? Like that's Oh, and it can go back and forth. Oh, that's that's really cool cuz a lot of the Figma stuff is kind of get it out of Figma, but it's hard to use AI or like cloud code to create designs in Figma. If I could use AI to create designs in Penpot, man, I'm I'm really sold then. I'm I'm excited. I'm selling myself on this thing. >> [laughter] >> Oh, come on. Hacker News is Hacker News is blocking me because I have an agent go and search everything. This like should work. It's But But honestly, the kind I do I do like controversy on Hacker News because I find that it is insightful. Um, I don't think the controversy here is really that big. Self-host it or use it, move on.
-
24:27
, obre el vídeo en una pestanya nova
Okay. Voice Box, remember earlier I said that there is another free dictation tool taking on Whisper Flow that's doing well. This is it. It's actually Voice Box. That's not the number one thing that they do, but if you look at their website, you'll you will see that they also are doing it. Well, the number one thing that they do though is they will clone voices. They will take text and turn it into beautiful, easy-to-edit voices. I think they've just been doing a phenomenal, phenomenal job here. This just keeps doing well. I actually, you know what? Um I I don't I don't like this you don't have to pay for 11 Labs and WhisperFlow subscription anymore. You never have to. This is This is just as good, but it's more than that. It's the ability to build it in, take control. Let me do a quick video here for you to show you what it looks like, and then I'll take your your feedback on it. >> This is what my real voice sounds like, and this is my AI cloned voice. Hi, this is Kevin's AI voice. You probably can't tell the difference. I was able to clone it in under 30 seconds. I will show you how to clone any voice on your PC with no subscriptions and no complicated setup. All you need is a short voice sample, and you can generate new speech >> It's fantastic. Um What do you What do you think of this one? Any any thoughts? >> Man, I love the local trend. Uh you know, and and voice does seem to be one of those things where the models aren't that big. I You might not have been paying attention, but I've been watching like as we look at some of these tools, uh how big like the peered one for example, or not peered. What What was the What was the other voice one we talked about here? >> Uh man, I forgot the name of it. >> I noticed that if you download Fluid voice, it's a 3 and 1/2 GB download. You're like, "Oh my god, that's a lot of space." Well, but it's because they have the whole voice model built into the app, and that's why it's so fast. That's why it's free. And then because it's local, you get all this control. And so these voice I think is a is a great beachhead into bringing, you know, whatever you're paying for service-wise online. Like I would disagree, Andrew. I think you probably could cancel 11 Labs. You probably could cancel WhisperFlow and replace it with uh you know, with either of the two tools that we've talked about today. >> Okay, fair point. I like the pushback on it. Um By the way, I've used AI generated voices here because you know me, Adam. I flub words all the freaking time. And uh I mean, we were looking earlier at the repos before we got started, and as I looked at the name, I mispronounced it or like reversed it. Anyway, I use it all the time. There's no other way. Otherwise, what am I going to do? Go and re-record? The re-recording is not going to sound as natural because when you use one of these, it kind of matches the rest of what you've put up there. >> That's right. >> Okay. Next. This is just a repeat of last week's system prompt leaks. You had a take on this, which is first of all, but let me let me just do this. I love these things. Look. >>
-
27:00
, obre el vídeo en una pestanya nova
Somebody leaked the system prompts of Cursor, Claude [snorts] Code, Lovable, and 30-plus AI coding. >> They really make it sound like it's this nefarious hacker who saved the day. It's It's not exactly that, as you explained. Would you do that again? >> Yeah, I mean, somebody leaked them. That's true. Like in the Anthropic case, Anthropic leaked them >> [laughter] >> live on their blog site. You know, it's not It's not so secret. But I would love a you know, I'm a sucker for for a dramatic music backing track. Yeah, all these guys, the the system prompt is the set of instructions that the model provider gives to the model for all the tools that you're familiar with using. Claude Code has a system prompt, and it goes something like "You are Claude Code. You are a helpful coding assistant. You work in this way. You don't work in this way." And every one of these tools that you've used has a system prompt. And and you know, the the leak is not There's no security risk here. There's no nothing. The The way that I look at these is I build a lot of AI tools, and what I've got in front of me are perfect examples from the best people in the world on how to write a system prompt. So, I'm looking at the template. I'm looking at the style, the formatting. I'm looking at Oh, for Opus, what did they include in the system prompt? And what did I think they would tell it that they didn't tell it? Because, you know, oversharing with a system prompt can also lead a a model astray. So, I'm looking at these, and when I'm writing a new system prompt, Andrew, I'm pointing Claude Code at a repository like this and saying, "Hey, go and learn from this real quick. I want to write a system prompt that, you know, Anthropic would say looks like it's uh among the best of the best." >> So, what you said is exactly what the Washington Post had said here. Use it to learn how to prompt properly. They also said you can see how they're dealing with things like biases. Like apparently Google has an anti-bias thing. >> Yeah. >> Um the one thing that's not going to work in a link that I have is this. What they did was they said, "Okay, see how with a different system prompt your your uh user prompt changes the results you get." And the reason it doesn't work is Washington Post is a paid subscription. And I wanted everyone who gets this report to have access to everything in here. And so I used the archive.is link for this one. I've been doing stuff like that before with these reports. We've had takedown requests in the past. People saying, "Wait a minute, the Washington Post, you have to pay for it. Make your users pay by linking." Um if we get it taken down, it's taken down. But for the time that it's up, it's in the description. Feel free to go and grab it, and then it's yours. Let's move on to the next one. >> [snorts] >> Uh
-
29:51
, obre el vídeo en una pestanya nova
we've also talked about this one before, Agent Reach. Give your agent the ability to actually read the internet, Twitter, Reddit, YouTube, GitHub by pasting one line. This is something that I incorporate into the work that I am doing here. Um basically what I did was I gave the this link to Claude Code. I said, "You know how I put this research together? What should we use?" And I said, "Well, I think you could use this, but not that. Let's keep going." And so that's where that's what uh that's what I ended up doing with this. Really helpful. >> That's awesome. [snorts] >> Um >> So, some of these some of these like AI scraping tools, like the AI browsers and even plugins for your browser, they're really good plugins that bring a sense of like agentic behavior to the human browsing experience. A tool like this kind of flips that on its head and it says, well, wait, you know, your agent really doesn't want to use a browser. It doesn't want to have to click all those buttons. It doesn't want to like sit as an extension. This is a way to give the agent direct access to all these sites on the internet and and they've got built-in, you know, there's there's this war of well, if I'm LinkedIn, I don't want you to scrape me. If I'm Twitter, I don't want you to scrape me. Well, these guys are doing their best to to stay ahead in that war where it's going to give your agent the best tools that work today for scraping LinkedIn posts, for scraping Twitter, for really any website on the internet. But they've got a couple of here that that they focus on to make sure that you're not going to get blocked. Like you just got blocked from Hacker News a minute ago. This tool exists to help your agent not get blocked in the same way. >> I guess I'm going to have to use this for Hacker News now. Oh. Um I do I do hate that I have to go through all that. All right, I've got a nice video here, but we've talked about this repo before, so I'll leave it in the report for anyone who wants it. Let's move on.
-
31:39
, obre el vídeo en una pestanya nova
Daily stock analysis, a free AI robot that reads the market for you for your watch list daily, and then it sends you buy sell hold a dashboard for your phones to help you make the right decision. We don't really cover much uh stock information, but I do like that people will go In the past, the the geeks used to have spreadsheets with what their stocks were doing. If they got a little bit more sophisticated, they might connect it up to something where they have real-time data. It's not for trying to um time the market. It's for trying to get a better understanding of what you have, the decisions that you make, and so on. And so now these tools are taking that same approach, and they're they're making it a lot smarter. Um that's that's where we are with this. I haven't tried it. I'm not I'm not into the stock market, but I want to identify that this is a big trend here, financial information on GitHub. >> [snorts] >> Yeah, I don't have much more to say than that. Okay. Uh and
-
32:42
, obre el vídeo en una pestanya nova
finally, number 10, OpenCut. CapCut is great, but I have to tell you the one thing that I do freak out about is giving CapCut the video editor access to my computer. In fact, I have it isolated somehow. And it's because it does some crazy freaking stuff. I don't like that when I use uh when I use What's the name of that company? When I use that company's software, um by chance, that it copies my clipboard. What do you need the clipboard copying? I don't like what it's doing on my computer. I don't like how I get to download it from the App Store, where I feel safer downloading CapCut, and then it sends me for the update to the web. And I was like, "No, the whole idea of using the App Store was to get around that." So, I do understand that. I don't think that I use all their features. And so, for the simplicity of editing, I understand that this CapCut alternative, called OpenCut, is uh it's easier. Uh sorry, it's it's desktop and it's in your control, and you can install it. I do want to give people just one big note about this one. It's not great. It's not a great video editing experience. It doesn't have all the features uh that CapCut has. It has had some problems where it's a little bit weirdly buggy. I don't think this is the right direction to go in. I think what you and I were looking at earlier, which is like let Claude code edit, I think is smarter. Um so, instead of duplicating CapCut, I like to see um open source take it to the next level and do what CapCut can't do. You know, go to Claude code, build it with my AI agents. Yeah, [snorts] I think that's a good point, Andrew, where where you know, I said this earlier when we were looking at uh at some of the other tools. It's like, well, my problem with something like Dscript or CapCut is that I'm limited by a kind of crummy UI. The AI is amazing. The generative capability is amazing. But sometimes I'm limited by the UI or just they don't have a button where I need a button to do something. And the open source version of CapCut is going to suffer from that times 10 because it hasn't been around as long. It hasn't had as much testing. And so the reason to go to to Claude code, open code, whatever and let those things use all these video tools is that you get the flexibility and you can let Claude code fill in the gaps. And I agree. I don't think this one is is going to do that. And this video is a hard problem. It's very complex. There's a lot of signal. The files are big. And so they've they've taken a pretty ambitious project on. You know, I I'm I'm excited if they can accomplish it because I would love to use this instead of CapCut like you said. But I I think it's a it's an uphill battle and they've got a long way to go. >> They are rewriting it from the ground up. They are going headless so that you don't need that that interface and your AI can use it. MC Hammer can touch >> goes. That might be amazing. >> Let me let me tell you something. If you go to new.opencut.app It has been down. It's been It's been down which doesn't give me a good feeling about it. I want them to succeed. Don't get me wrong. >>
-
35:24
, obre el vídeo en una pestanya nova
Usually Andrew, if I'm going to build around an app like this and I get a couple bad signals early on, you know, it's unreliable, the website's down. The two things that I go and look at. Go back to the GitHub for me real quick. I I want everybody to see this. This is like you know, I I don't want to learn this tool if the tool's going to go away, right? So number one, scroll all the way down. Who's sponsoring this? So foul.ai, okay. >> a good sponsor to be supported. They they're great at video. Uh >> But but I I I want to see a reliable sponsor, but I also want to see multiple sponsors. Because why is foul.ai sponsoring this? And you know, if in any place where you get one sponsor that's paying all of it, well, they're kind [snorts] of in control and they're going to make decisions that are best for them. I want to see a couple competitive sponsors that all make sense there. And then then the next thing is I scroll all the way up and look at the files here. When were they updated? 5 days ago? That's a pretty good signal, right? You know, some of these you'll look at it and it'll say last month or 3 months ago. I'm looking right here on the right side. 5 days ago, 5 days ago, 3 months. And and if you're not seeing a lot of activity there, and it's not very reliable, ooh, I cut and run. But here, like one sponsor that's pretty good, I'd like to see more, uh but a lot of activity, like they very well could could come out of this slump. >> I'd like to see them do well. I like to see more video editing. Um I do have again in the report, there's a tweet that they put out where they were talking about how they're rewriting it, what their vision is, they're taking people's uh requests to heart, like the keyboard first mode, etc. Okay. >> Cool. >>
-
37:21
, obre el vídeo en una pestanya nova
I've spent hours on this report, literally hours. Even though I have obviously Claude Code helping me put this thing together, I spent hours because I want to make sure that there's no hanky-panky, that I don't end up with somebody who is clearly trying to game the system and get in here. I want to make sure that we've got useful information, that I'm curating the videos to make sure that they're helpful. If uh if this has been helpful, I would really appreciate it if you gave us a like, if you did subscribe, or even frankly if in the comments you told me what you hated about it, because that helps me improve. I want this to be the best uh way for you all to understand what's going on in open source, and more importantly, how you could use it and build with it. Um and Adam, I was doing this last night with my with my um >> [snorts] >> Michelob Ultra, which is like the wussiest beer. That's what I'm down to, but I'm trying to like watch my calorie intake, and at the same time I feel like Thursday night, baby, everyone's watching soccer. I'm sitting here at my desk just preparing for tomorrow. Let's enjoy a little beer. >> Me and AI, yeah. Yeah, you've had kind of a week like that, too, where it like you couldn't take much time because you had kid duty, right? >> gosh. I Yeah, I've got my my kids alone this week. My wife's out of town, and and so I'm I'm off work basically at 10:00 p.m. And yeah, with one or two nights I think I sat down and had a beer, chilled on the couch, and was like, man, I got to crash. I got to bed. But yeah, it's you know, you you need a little refreshing. And honestly, Andrew, like AI it's exhausting. It it I'm I'm more tired at the end of my work day when I've worked with AI more. It's just constant decisions, constant back and forth, constant frustration. I I I've not cursed at my computer so much in years than like compared to the last 6 months. But it's also so fun. It's so it's it's addicting. And I But it man, it it drains me. I put so much so much into the apps that we're building, so much into the shows like this that we're working on and everything that feel like at the end of the day I'm just done. >> I'm glad to hear that you're frustrated. Sometimes I feel it's cuz I'm stupid or I'm missing something and a smarter and more engineering-oriented person wouldn't have that problem.
-
39:27
, obre el vídeo en una pestanya nova
Okay, now that this video is over. So far what we've talked to you about is GitHub. I now have actual apps that you can use right now that are AI-based that we hunted down for this week. There's a link for you to watch it right now. Just go use it. You don't have to incorporate it. You don't have to do anything. It is amazing what you could achieve or frankly just sit back, watch it with a Michelob Ultra, and watch your weight like Andrew. All right, go see you in the I'll see you in the next video. >> Bye.