maddada • ghostex.dev

Ghostex 10 is out! Put in a TON of effort! Would love your thoughts 😄

Ghostex 10.47.0 is out! Hundreds of improvements in the past 2 weeks but I never post 😅 Please check it out if you want to try the Rust based ADE with the best way to work with Coordinator agents out there hands down. You'll never go back to managing threads after trying Coordinators. • Much nicer inline HTML questions/docs template. • Coordinators spin up workers on multiple machines more often now • Added a Coordinator work board (Shows all your high-level work statuses as a list/board) • One-click answers in chat (shown in screenshot) • Optimized resources usage even lower than before • iPhone voice mode (TestFlight) Full Changelog: github.com/maddada/Ghostex/r…

Damn, Anthropic needs to fix this ASAP. @ghostex_dev is not affected at all since it's NOT based on ACP or agent-sdks. All your threads (even gui ones) are running on real Claude Code CLIs in Native Ghostty terminals unlike 90% of the other ADEs out there. Native Rust & GPUI.

IMO, the best way to use Claude and Codex in the same repo and share context between them is Ghostex Coordinator threads since the main agent has all the info and can spin up any other harness for sub-threads.

You'll trim this sidebar so much in a week or 2 when you switch from a thread per task to a few orchestrators managing all your work. Managing threads is mostly a waste of time. Trust. Let 1 smart agent manage all your sub-threads, it's so freeing. (Look into @kunchenguid's firstmate) Ghostex.dev has been doing this for a while now and my sidebar is so empty now compared to before, while I'm doing way more than before.

This is the way, Ghostex does the same since I added chat. Don't understand why others just attach the image to the whole message. Makes it much harder to reference a specific screenshot.

I do almost ALL my work in Ghostex this way (1/2 orchestrator threads passing work to sub-threads). @kunchenguid was hyping it with Firstmate so I went and implemented great UX on top of this: 1- Non-blocking questions UI that you can answer whenever you're available 2- Ability for sub-agents to be threads in any other agent 3- Skill for in-chat html mockups that let you pick variations by clicking on the design you like (sent to agent when done) 4- A new handoff format for these Orchestrator sessions (Done, Kicked off, In Progress, Needs you, Untested, etc.) 5- Links in the chat that take you to previous async questions you didn't answer.

Love this "new to me" feature in Claude Code! It automatically compacts context when you go idle and cache is still alive. This can save a lot of tokens over time. Hope the others copy it.

I don't like it, but what @robinebers is saying is true. I cancelled 2 Claude accounts when Astra came out then came right back with 3 accounts after trying it 😂 Still not "happy" about 50% fable usage, their comms, Opus 5 being absolutely useless, and many other things. It's business in the end, they're free to charge what they want, and I'm free to jump ship when someone leap frogs or matches them (like before fable) I prefer OpenAI as a company (they feel more generous and chill, codex cli is open source, resets), but Anthropic have the most worry-free/enjoyable model to build with 100%. I would just use fable 24/7 if I could. The solution: (This is just a shpeel about using an ADE that supports all providers and subs, skip if not interested :) This situation is one of the main reasons why I don't use any first party apps and made my own ADE instead which works with all the well-known Agent CLIs. This way I can't be locked-in no matter what. All my convos are mine and in 1 place, and I can handover between agents with 1 click. I even put a standard GUI over the CLIs so all of them have the exact same hotkeys and flows. No more different hotkeys/commands between CLIs. If Anthropic go through with limiting their SDK usage then my app would still work since there's a real terminal running for each session 🧠 All the chat view is doing is showing you the text from the terminal in a nicer way + forwarding your input from an easy to use composer to the terminal. You can jump between chat/terminal view anytime too. iOS & Android working perfect :) Such a fun project. (Browser, kanban, automate, etc. are all fully optional extensions, keep it minimal if you want or enable them / create your own extensions if you want) Thanks for coming to my TED talk and try Ghostex if you haven't!

GPT 6 Astra is generating atrocious UI for me for the 10th time. It's good at web design if you put in the effort, people have posted the sample websites and it's great. But it's insane how bad it often is when it comes to app UI/UX and general styling. @robinebers Not sure if this is what you mean about the model making you rage but omg I can't give this model anything that has to do with user experience or anything visual at all. I added a handover feature to my ADE just so i can pass the thread's user/agent messages to fable and continue from there.

I claimed @maddada on OpenShaders and got my one-of-one shader. What will yours look like? openshaders.com/@maddada Awesome site @davidhaz!

Check it out! I think I made the best model/effort picker out there. Option + p for both. Matches Claude's. Also in this demo I show how Ghostex combines the great UX of Chat + the power of Agent CLIs. It's truly the best of both worlds. Switch instantly anytime with ⌥/alt + g. Based on Ghostty, GPUI, and Rust. Not using electron/tauri. All of this solo. 9.0 coming today: ∙ Full customization + extensions system ∙ Make it as minimal or feature rich as you want ∙ Multi-account support + auto-switching ∙ Auto retry & continue on errors ∙ See sub-agents transcript in Claude & Codex ∙ Queueing messages for Claude Code ∙ Handover sessions between any 2 agents ∙ Any agent/harness can control any other harness. ∙ Much improved Agent CLI chat view ∙ Improved performance & reduced ram use ∙ Windows x64/arm & Linux support (arch/deb/rpm) - need contributors to make it perfect. Please let me know what you think. Good/bad feedback is welcome :)