Hunter
255 posts

Hunter
@snappyhunter
just building stuff. prev founding eng: @auto_gpt. prev eng: @snapchat.
Los Angeles, CA Katılım Kasım 2021
271 Takip Edilen535 Takipçiler

Claude Remote Control is extremely nice
Can edit on MacOS or iOS in Claude app on my production server from anywhere
A bit more friendly to the eyes than SSH

@levelsio@levelsio
My new command for Claude with remote control on yolo mode: c() { IS_SANDBOX=1 claude rc --dangerously-skip-permissions "$@"; }
English

You can now connect Slack to Claude on Pro and Max plans.
Search your workspace channels, prep for meetings, and send messages back to keep work moving forward—without leaving your conversation with Claude.
Get started: claude.com/connectors/sla…
English

@alxfazio @flydotio @Zai_org Love this idea. I started tinkering with Clawdbot after using CC since release. Decided it would be better to build a tailored solution for my needs. Would love to see some design/architecture document for some inspiration when you are ready to share. Dropped a follow
English

i'm building a fleet of vertical agents where each one gets its own full computer (100gb persistent disk, 8gb ram, ~600ms wake from hibernation) using fly.io sprites and z.ai glm 4.7. they're built entirely on claude code cli, no openclaw and no claude agent sdk. `claude --print` (non-interactive mode) is what makes this possible. the hatcher sends a prompt, claude code runs headlessly on the sprite, and streams structured json back. you can talk to them from discord or drop into a terminal and pick up the same session from anywhere
we call them droids. fly.io sprites give them ~$0 compute when idle and auto-hibernate after 30 seconds of inactivity. when a user sends a message, the sprite wakes in ~600ms and picks up right where it left off. they'll be hatchable from a base config using what i call a "droid hatcher," similar to botfather on telegram
we were already building useful vertical agents with claude code, but we're all busy and travel a lot. having to sit down in front of a laptop and open a terminal just to use claude code (because the app and claude.ai don't have the same toolset) was creating a lot of friction. creating a new agent was an entire project. sharing one meant pushing to github, sharing the repo, pulling updates. group chats with claude code didn't exist, you couldn't have multiple people talking to the same agent in isolated threads. with this setup anyone in the company can hatch a new droid directly from discord using the droid hatcher, share it with others, or co-work on one together, all in full safety
i wanted something like openclaw for my business, but openclaw is too opinionated. it abstracts away too much of the claude code configuration, which kills customisation and portability. it would also force me to rewrite the claude code setups we already share across employees. and the agent sdk is too high-level for what i need. i want each droid to own its entire machine, not run inside someone else's runtime. because these are full claude code instances on isolated vms, the agents can self-improve, self-edit their own CLAUDE.md, and their machines are fully auditable. we can ssh in and inspect anything at any time
this system solves all three problems. each droid is its own machine (no git pull needed), lives in a discord forum channel where each post is an independent session, and multiple people can interact with it concurrently
i can already resume a droid's session from a terminal whenever i want, instead of going through discord. their main home will be a dedicated discord server. i chose discord because i used midjourney a lot and i know what's possible on discord in terms of ux for bot-driven workflows. it also avoids dumping noise into our main slack channel, but they'll still be mentionable from slack when needed
anyone in my company will be able to create and customise their agent, all from a single chat. new people, collaborators, and contractors can be onboarded safely in discord, since each droid lives in its own forum channel and permissions are managed directly through discord roles. no complex business logic needed. contractors can even bring their own api key and hand it to the droid hatcher to spawn a droid they pay for, while still working transparently inside our org
the system supports third-party model providers like z.ai for tasks that don't need frontier intelligence. glm 4.7 is fast and cheap enough for most vertical agents. this also matters because anthropic has been banning accounts that use their max subscription inside third-party wrappers, and running through a separate api provider means your anthropic account stays clean
i think this is a better and safer version of both claude cowork and openclaw. each droid has its own isolated vm. not a shared sandbox, not a container, a full machine. if something goes wrong, you restore the entire machine from a previous checkpoint. no need to buy a mac mini just to feel safe, and no risk of an agent messing with your actual computer
this is probably the most fun llm project i've worked on so far. right now i only have a working slack prototype, but the signs are all there, and it has me genuinely excited. the full system with discord forum channels, the droid hatcher, and session forking is what's coming next. i already have a full design spec for the discord system and i'll share it when the time is right. i'll keep posting updates as it takes shape
if i could, i'd invest in fly.io, because they've built an incredible solution for what i think is the future of agents. the full comparison of infra alternatives that led to this choice is in the post below



English
Hunter retweetledi

@Asmongold Praying for you. It gets a little easier everyday. Lost my dad 2 years ago
English
Hunter retweetledi


Juggling 6 projects now for a few different clients. Thank you conductor.build @charlieholtz and #claudecode @AnthropicAI
English









