thanks for using claudable before! for server costs, the sandbox previews have auto-stop functionality so the actual runtime is pretty short. keeps costs manageable.
for deployments, cloudflare static hosting is basically free, and we only charge a small fee for custom domains since that's where we have actual costs. the beauty of leveraging existing platforms is that we don't have to run expensive infrastructure ourselves. most of the heavy lifting (the actual coding) happens on your CLI agent subscription, not our servers!
Oh I really like your product! you're right about how it works. the desktop app sends oauth tokens to connect to the cloud infrastructure.
dyad looks really cool, love the local-first approach. it's exciting to see so many different takes on this space right now. what are you focusing on with dyad these days?
Thanks for reaching out about the hosting platform! really interested in the European market, will definitely drop you an email once we get things more organized on our end.
You make a great point about the messaging. claude code is definitely comfortable for many devs right now. but we're seeing people constantly switch between codex, claude, and now newer agents like droid and glm. we're positioning ourselves as a mixture of agents platform where you can leverage the best tool for each task.
also yeah, CLI agents are amazing for coding but not great for instant previews or deployments. that's exactly the gap we're filling. the positioning will probably evolve as we learn more from users, but that's the core idea: bring your favorite CLI agents, get the missing pieces (preview + deploy) without paying for another full coding platform.
appreciate the feedback on making this clearer upfront!
ah, I think there's some confusion with the open source version! The OSS Claudable runs everything locally on your machine. This new Clink product runs in the cloud with Daytona sandboxes.
sorry for the confusion, we should clarify that better. the 'locally' copy was from the original OSS project description. will update to be clearer about the cloud infrastructure for Clink!
Most app builders use lightweight coding agents and infrastructure. we're leveraging the full power of CLI agents (Claude Code, Codex, etc) as the engine, which means you can build with any frontend stack like react/next/svelte and any backend like fastapi/nestjs/flask/ruby, then deploy everything as a container. this is live in beta now.
also, if you have an existing project, say a python ML model, you can import it and add a frontend to create a complete deployable service. the CLI agents are just way more capable than typical app builder agents, and we're giving them the infrastructure to match that capability.
Good question! while we use CLI agents as the coding backbone, our focus is really on the application building experience: live preview and seamless deployment.
The key difference is that we give you an isolated sandbox container where you can see your app running instantly, then deploy it with one click. so instead of just getting code output, you get a runnable application that you can interact with and ship immediately. openCode is great for pure coding tasks, but we're optimized for the full build-preview-deploy cycle.
Ah, Vercel is what we use in the open source Claudable version! For Clink, we currently deploy static/frontend to Cloudflare Pages and full stack apps to Railway, both work great for different use cases.
We're definitely interested in adding more deployment targets. What hosting platforms would you prefer to see? we can prioritize based on what the community needs most
For now, we support switching between models within the same provider, but we're currently adding the ability to switch providers mid-session. We're working on agent switching similar to Cursor's LLM switch so you can seamlessly move between Claude Code, Codex, GLM, etc. during development.
We're also exploring auto-switching based on task type, our model would automatically select the best agent for each task. For example, using Opus for frontend work and Codex for backend refactoring. The goal is to leverage each agent's strengths without you having to think about it.
We're using Daytona for development environments, which gives us proper isolation out of the box. Each build runs in its own sandbox workspace with full container isolation, so there's no cross-contamination between projects or users.
What I liked here is how unglamorous it is: tiny prompts, context only when needed, evaluate against real usage, and resist multi‑agent stuff until a single boring pipeline is stable. They also pair rule checks with model checks and expect some reward‑hacking. Curious how others keep evals from drifting in production.
Solid, practical write‑up on running an “agent” in production. Biggest takeaways for me: keep the base prompt tiny, inject context only when it’s actually needed, and eval on real traffic before adding multi‑agent complexity. Has anyone tried that “just‑in‑time” instruction style at scale—what broke first?
Love the demos, but the bar for “digital restoration” should be provenance + uncertainty-by-default: clearly label what’s scanned fact vs. modelled conjecture, keep raw data reversible, and publish model/data cards so future scholars can audit the guesswork. Without that, we’re at risk of preserving our present biases more faithfully than the past.