twobrainy

Looks very cool.

I've also built a daemon->control plane system built on ACP and Pi RCP/Something like (but not as good as) Pi Durable, won't share here as its your post.

I have a slightly different twist on the organisation, and UX of project and tasks but the back end is fundamentally the same: Let users run tasks on their laptop, "AI Shed" server as DHH calls it or whatever.

Then a UI to manage those tasks per-repo. My attempt also manages worktrees, merging to main, pre-commit checks and PR creation as I believe yours does.

Not aiming to be dispiriting for you OR me, but I wonder where the space is to try and build paid services over these primitives, especially when code is so cheap now, and Claude is launching things like "Claude Mods" so users can modify the UI of the "Big LLM" harnesses.

Obligatory - everyone should try Pi. I've been using it since I came across it via the excellent https://stencil.so/blog/the-harness-problem blog in December.

After being a Claude-hater for the last few months, frustrated with weird verbosity and slow churning response, not to mention dealing with a full-organisations worth of shared skills and giant CLAUDE.md docs slowing everything down - super-light Pi is a delight.

Even still, I normally delete every package, skill and agent.md and start configuring again when a new class of model (Astra/Fable/5.5) comes out - I get amazing results with Kimi K3 in Pi. I really really love pi lol.

Other than quite liking Codex desktop now... I also recommend everyone tries building a harness of some sort. Vercel Eve is really fun to play with if you have a Vercel account and consider that a type of harness, or at the very least build a control-plane/UI for a harness on top of the Pi RPC SDK which is incredible.

I personally have a custom web ui powered by diffs.com and drag and drop "pipelines", where I can orchestrate steps instead of skills (although they're basically skills) + custom compaction and summary relay between the steps with a small fast model.

I also have iOS/Android apps that look like Trello boards, which communicate over Cloudflared/Tailscale to a thin Pi server that provides github integration and some custom tooling to render "ask user" prompts and co-ordinates worktrees and PR creation.

https://fx.sh is also cool.

All in all - use your spare tokens to build and try harnesses, build alternate harness UIs, and build alternate workflow co-ordinators away from just "/invoke-skill" or a TUI-based chat.

Quite funny because Earendil is the old english word for Lucifer aka... Satan

Interesting tool. I've seen this done with an agent and google workspace CLI.

I like the idea of being able to bulk assign labels before a cutoff and the other Gmail API driven stuff, but the AI work is a bit unclear.

"Some message metadata and snippets", what exactly does that entail? Partial message bodies?

Handing over email access, however limited, to a 3rd party wigs me out. Would you consider a runnable local tool instead of a hosted service? Or a bring-your-own-keys system??

I realise people are pretty cavalier and integrating their Google Accounts with ChatGPT desktop anyway, so maybe it doesn't bother many people.

Does your OpenAI API account have zero data retention mode?

I pay money so I can sleep easier at night!

Envoyer and Forge for PHP stuff.

Render/Heroku for node.js stuff. Netlify for static stuff.

For the more complex bits I have, I use Pulumi to manage everything through IaC. It's still complex but at least it's robust and I can sense-chech changes to infra through a PR to myself.