Maybe it's time scrapers actually paid publishers via something like HTTP 402 for their data instead of an arms race with Cloudflare on one side and residential proxies on the other.
One feature suggestion: webhook support for row changes:
If my Sheet updates (say, a new waitlist signup), I'd want to trigger for e.g. a Slack notification. Supabase has something similar with their database webhooks. I use that extensively for kicking off signup workflows.
Can confirm good canvas renderer performance. Just tested it with a real-time smart meter dashboard — 62 meters streaming over MQTT via JustinX.ai (our data ingestion platform), 60 msg/s (peak), 500ms refresh. Almost no flicker, smooth updates across all meter cards. Much nicer way to handle high-frequency streaming data than xterm.js.
I think the A2A space is wide open. Great to see this approach using App Server and Channels.
I tried built something similar (at a high level) for a more B2C use case for OpenClaw https://github.com/agentlink-dev/agentlink users. Currently I think the major Agents have not fully owned the "wake the Agent" use case fully.
Regardless this is a very cool approach. All the best.
I think this would be super useful as a claude code / visual studio extension, so I can see the cost impact of changes I am making in code. For example by adding a prompt pre-processing function how much am I actually going to save in LLM token cost.
Very nice idea. I have sent a little note to me and my founder in the future. It has a nice "time capsule" like feel to send a message out. Really makes you think about what you want to say.
There is a real risk but probably not directly from someone targeting you. Your agent reading a webpage or email that happens to contain injected instructions is a risk. It is really a surface area problem. I would suggest you ask claude/whatever to scan your OC dirs regularly.
Cool project — the "agent is a folder" philosophy is genuinely appealing. I spend most of my day in Claude Code, which is basically a primitive version with: flat memory files, file tools for self-management, context that gets trimmed.
The one thing I'd push on: the bet that the agent will reliably manage its own memory with read/write tools hasn't played out for me in practice. Claude has file tools and a memory directory today, and it still forgets things I've told it dozens of times — the bottleneck isn't storage, it's that the LLM doesn't reliably decide what to save or when to retrieve. That said, preserving full JSONL history on disk and only trimming at inference time is a genuinely better model than lossy compression — I wish Claude Code did that instead of auto-compact.
Have you thought about layering lightweight semantic retrieval over the knowledge/folder so the agent doesn't have to manually grep its own brain?