Chats 2.0
From a chat app to an agent platform that gets work done 2.0 is about agents: a persistent place for the AI to work, a way to publish what it builds, and collaboration that feels like working with a reliable colleague.
Persistent Docker workspaces Containers outlive a single chat — the environment you set up last time is still there next time.
Long-lived containers alongside today’s per-session ones Dedicated volumes, so databases don’t live inside the container Per-chat choice of which containers the AI may see Quotas on CPU, memory, network and container count; admins unlimited One console: users see theirs, admins see everyone’s Ship a website from the chat Tell the AI “make it public on this subdomain” — and it’s live.
Containers serve the site; traffic enters through the gateway Caddy 2 gateway and port mapping configured over its API Pluggable DNS, starting with Alibaba Cloud DNS and Cloudflare An agent that asks — and listens It asks when things are unclear, and you can chime in mid-run.
Ask Question: several clarifying questions at once; the ReAct loop resumes on answer Plan / clarify mode must ask about unknowns; normal mode aims to finish in one go Steering: no need to stop — your note is injected after the next tool call Upload an image once; later turns reference the provider’s file ID.
Uploads go to our own storage and to providers that support file IDs No more re-sending Base64 or URLs — much faster multi-turn image chats Across Responses, Anthropic Messages and Chat Completions, with expiry handled Streams that survive a refresh Reload the page and the AI is still visibly at work — and still stoppable.
SSE progress persisted to the database and replayed on reload Stop tokens persisted too, so Stop works after a refresh Cleaned up periodically once a turn completes Watching: subagents. Interesting, but speed and control come first — not part of 2.0.
Directions, not commitments — scope and timing may change.