Benoît Devilliers

@petergyang Added this to my openclaw and memory is much better github.com/martian-engine…

My personal experience on the drawbacks of using Claude Code vs. OpenClaw as a personal assistant from my phone: Claude Code - Doesn't have dangerously skip permissions via remote control. Also remote control doesn't feel super reliable - Doesn't have voice replies so doesn't feel as personal OpenClaw - Tends to forget things randomly still

@bestbubbledev VPS for now, haven't ran into limitations yet, will invest later if I reach some. I prefer starting small

Where do you run your OpenClaw and why? Mac Mini vs Cloud

@ToolsOttho @nicolasbucaille C'est des sub agent mais qui peuvent utiliser Claude ou Codex en mode CLI

@benoit_dev_ @nicolasbucaille Ok chaque agent est un openclaw ou ce sont des sub agent ? Je n’ai jamais tenté openclaw en mode dev, j’avoue que Claude me suffit

figma make est devenu hors de prix... 5 prompts pour fumer 1000 crédits sur sonnet4.6. => Blague Un meilleur outil à me conseiller ?

@elwatto I'm building a dedicated UI, you can chat directly from here, I have push notifications to my phone. I have projects/Kanban connected to linear. The idea is to have visibility on what agents are doing

Getting diminishing returns from having OpenClaw run all subagent conversations in a single Telegram thread. How are people splitting this up? Not a fan of Telegram topics… would a Discord server be a better workflow?

@0xDesigner I think the best UI is always to have optionality, full chat isn't the future, it's a mix of chat AND UI

no one is talking about how the best part of paperclip is giving claude code/codex a visual interface. forget orchestration and ai companies for a sec. that you can build a visual interface to escape the chat thread hell is mind blowing. i’m tired of reading a novel everyone i want to make a small change. it’s nauseating.

@ToolsOttho @nicolasbucaille Une UI pour mon Openclaw

@benoit_dev_ @nicolasbucaille Attends, c’est quoi ?

@nicolasbucaille @ToolsOttho Oui pour explorer ça va vraiment vite

@benoit_dev_ @ToolsOttho Le design, mon gros point noir, est maintenant réglé grâce à cet outil. C'est loin d'être parfait, pas aussi bon que l'humain, mais ça permet de faire de sacrés maquettes quand même.

@ToolsOttho @nicolasbucaille Ca fait plusieurs fois que j'en entends parler, j'ai testé ce matin et c'est une dinguerie !

@nicolasbucaille Ho wow, merci !!!

Does anyone use PR environments? It sounds ideal in theory, have your agent work on a PR, get an isolated database/frontend to check the changes. test and validate. Is someone doing it with @supabase? I found the integration clanky and my CI isn't reliable just yet

@T_Zahil @cursor_ai What's your take on asynchronous agent work that are proactive? Ie waking up to a bunch of PRs. As far as I'm concerned there are two modes: - Fully focused on making things work and I'm in the IDE - Making small tweaks and I don't need to be there

We’ve gotten used to “slow” models: multitasking, spinning up agents, juggling contexts… But the future is super-fast AI. One chat, one task at a time. No context switching. Full focus. Composer 2 @cursor_ai is giving us a preview of this, I love it

When generating elements with AI in apps there are a lot of opportunities to create some cool waiting screens... what are the bests you have seen so far?

@InsomniacPM22 @elvissun can you elaborate? is this format readable by humans too?

@elvissun @benoit_dev_ Indeed. Though i have been able to create an agent native api format they like even more!! I know it sounds really crazy that’s some random guy did it but yeah, I benchmarked it.

this is the future of saas: one backend with two front doors. one for humans. one for their agents. same building. your users click buttons. their agents call your api. both do the same things. why: 1. agents are the new power users. people are already wiring claude code, codex, and openclaw into their workflows. if your product doesn't have an api, you're invisible to them. 2. no drift. one canonical api layer means one source of truth. ui and api never diverge. you must support agent to support human. 3. skills is the new distribution. when your api works with any agent, agents find their way in themselves. your product becomes a tool for their agent, not a tab a human have to visit. build for both from the start.

@morganlinton Have you checked @composio? it's a good way to manage credentials

I think credential isolation is going to be an increasingly hot topic in the AI agent world. Can't say I feel like it's being done in the most robust way across-the-board. Honestly, in the current phase we're in, it really us up to all of us running and building ai agents to make sure we really understand security and things like credential isolation, and make sure it's being done right, don't just trust that a third-party is.

@yelkhayami could be a ghost cat

@benoit_dev_ next time inshallah

two clicks and you're a ghost

@hemeon I have been working on this topic recently, best approach seems to have a good template to begin with :)

I'm manually designing a deck this morning and it's taking me about 2 hours by hand. Spent all afternoon yesterday trying to make it with AI and wasted way more time typing than just making . Don't forget you can just design stuff with your mouse.

@julianlehr I think we will end up in an in between for you have UI elements living in the chat. MCP apps were a step toward this modelcontextprotocol.io/extensions/app…

A hill I'll die on: Current LLM chat interfaces are a regression from GUIs. Actions that used to be links, buttons, or keyboard shortcuts are now things I have to spell out in conversation. Why?

@larsencc Use claude code from cursor, browser management is just too nice (free plan works like a charm)

Should I just delete Cursor at this point? I find myself using Claude Code 95% of the time...

@kentcdodds I think we must define product principle and add constraints, since we can build anything these days

Here's how I'm trying to figure out the skills that are durable enough to not be replaced by an agent... Imagine a future where agents are able to manage an infinite amount of context and have access to all of the same tools that you have access to. In that world, what unique value proposition does a human still have over an agent?