kevinlee

building Codync

[kevinlee @lbki34064963]

@Ga_Vasques Thanks for putting Codync on tier S!

[kevinlee @lbki34064963]

Codync is now on Windows.

Install it, click it in the tray, scan the QR code with your iPhone. The Claude Code, Codex and Cursor bots on your PC are one message away, and you approve their changes in one tap.

Free and open source.

[kevinlee @lbki34064963]

Windows: https://t.co/16Y1jrPzGU
(not code-signed yet, so SmartScreen asks once: More info → Run anyway)

iPhone: https://t.co/lk8q5trCRj
Code: https://t.co/peny3P5D4d

[kevinlee @lbki34064963]

Codync is now on Windows.

Install it, click it in the tray, scan the QR code with your iPhone. The Claude Code, Codex and Cursor bots on your PC are one message away, and you approve their changes in one tap.

Free and open source.

[kevinlee @lbki34064963]

One agent runtime, anywhere WebAssembly runs.

I compiled the real Pi Durable engine into QuickJS WASM. Same engine runs in the browser, Node, and on-device inside iOS and Android apps.

Kill it mid-run, reopen, and it resumes from the last checkpoint.

MIT. Link in the reply.

[kevinlee @lbki34064963]

https://t.co/HiDKmrNu6c

The whole point is the runtime. Pi's engine lives inside one WASM guest, and the host only supplies three things: a model transport, a durable journal, and the tools you grant it. Swap the host and the agent doesn't change.

Where it runs today:
Browser: Web Worker + IndexedDB
Node: macOS, Windows, Linux
iOS: WKWebView, calling a real Swift tool
Android: WebView, calling a real Java tool

It's the actual upstream Pi scheduler, not a simplified loop, so steering, follow-ups, fork, compaction, extensions and subagents all come along. API keys stay on the host and never enter the guest.

Easiest way to see it: open the browser playground, click Run, hit "Interrupt runtime" (it really kills the Worker), then "Reopen & resume".

Still experimental (v0.2). Install from the GitHub release, not on npm yet.

Inspired by Finn Voorhees' PiDurableKit, which does this on Apple with JavaScriptCore. I wanted the same thing anywhere WASM runs.

[kevinlee @lbki34064963]

One agent runtime, anywhere WebAssembly runs.

I compiled the real Pi Durable engine into QuickJS WASM. Same engine runs in the browser, Node, and on-device inside iOS and Android apps.

Kill it mid-run, reopen, and it resumes from the last checkpoint.

MIT. Link in the reply.

[kevinlee @lbki34064963]

After looking at a lot of agent benchmarks, my takeaway is that the harness has less impact on performance than I expected. The underlying model still seems to do most of the heavy lifting.

When a model performs especially well in its own provider's harness, I suspect some of that advantage comes from the model being trained or optimized for that environment.

Take Claude + Claude Code: if Claude is trained on trajectories or transcripts from Claude Code, the resulting gains would come partly from improvements to the model itself. They wouldn't necessarily show that the harness alone makes any model much better.

To me, the clearest value of a harness is in the tools, features, and workflows it provides. Those matter. But for raw task-solving ability, I still think the biggest factor is how capable the underlying model is.

[kevinlee @lbki34064963]

@yifanxu_ephai 大佬厲害 目前也想要往 agent infra 探索發展

[kevinlee @lbki34064963]

Codync is now on the App Store 🚀
A free, open-source alternative to Grok Bot / Muse / Dots.

Run Claude Code, Codex & 40+ agents on your own computer and message them from your iPhone.

⭐ Star us: https://t.co/peny3P5D4d

[kevinlee @lbki34064963]

Codync is now on the App Store 🚀
A free, open-source alternative to Grok Bot / Muse / Dots.

Run Claude Code, Codex & 40+ agents on your own computer and message them from your iPhone.

⭐ Star us: https://t.co/peny3P5D4d

[kevinlee @lbki34064963]

LCU by @0xpolarzero (@amontlabs) takes Codex's computer use out of the Codex app and into other agent harnesses. Great idea.

I wanted it as a plain Claude Code plugin, so I built Cluadex: Mac only, ~700 lines, 2 commands, Liquid Glass approval panel.

https://t.co/hPo5N2XCcx

[kevinlee @lbki34064963]

Cluadex is an independent reimplementation, not a fork, and shares no code with LCU. The idea of reusing the installed runtime, and the host guard, both come from LCU.

Install:
claude plugin marketplace add leepokai/cluadex
claude plugin install cluadex@cluadex

Needs an Apple Silicon Mac with the ChatGPT desktop app and Computer Use already working in Codex. If you want Linux, Chrome control or another harness, use LCU: https://t.co/vdhN5j9Mwc

Unofficial, not affiliated with OpenAI or Anthropic.

[kevinlee @lbki34064963]

LCU by @0xpolarzero (@amontlabs) takes Codex's computer use out of the Codex app and into other agent harnesses. Great idea.

I wanted it as a plain Claude Code plugin, so I built Cluadex: Mac only, ~700 lines, 2 commands, Liquid Glass approval panel.

https://t.co/hPo5N2XCcx

[kevinlee @lbki34064963]

@yetone 我最近也使用 WebRTC + cloudfare turn 实作了一个
已经上架 App Store 可以玩玩看喔～

https://t.co/peny3P5D4d

[kevinlee @lbki34064963]

@yifanxu_ephai @MingDynastyVase 最近也做了个类似的 open source 可以参考一下

https://t.co/peny3P5D4d

[kevinlee @lbki34064963]

Really happy to see Pi Durable officially (@pidotdev) launch.

The original idea behind nimplex came from a problem I kept thinking about: how should cloud agents reliably recover when something goes wrong in the middle of a long-running task?

Processes crash, sandboxes die, networks disconnect — but the agent shouldn’t have to start over.

That led me to explore durable execution, persisted state, and resumable tool calls in nimplex.

Seeing Pi Durable now tackle many of the same problems is incredibly exciting. It feels great to see an idea I cared about becoming a first-class primitive in Pi.

Long-running agents need durability.

https://t.co/4ErbPjZCqv

[kevinlee @lbki34064963]

@ZouhairMaj It’s a very good’s idea!
I will put this feature in my future plan.

[kevinlee @lbki34064963]

I liked Grok Bot's bot-based UI so much that I spent 5 days building an open-source alternative.

1 person + Opus 5.5. That was the whole team. Still feels insane to me.

Almost any coding agent, on your own Mac or Linux, built in Rust. Free.

Link in the reply.

[Zouhair Majzoub @ZouhairMaj]

Open-source alternative to Grok Bot/Muse/Dots

[kevinlee @lbki34064963]

@lidangzzz 之前也是考虑到隐私的问题
不敢给 grok bot 太多我的 credential
所以做了个本地 grok bot

https://t.co/peny3P5D4d