← sources

Connecting Claude CLI to antigravity IDE for a weird mobile workflow

source: reddit (reddit:ClaudeCode)published 2026-11-02 21:21ingested 2026-11-03 02:42

retained text

So I had this idea a few days ago, started building it, and halfway through it basically inverted itself. Wanted to throw the current setup out here to see if I’ve built something genuinely useful or if I'm just deep in my own over-engineered bunnle. Could be shot, if so, just say!

**The initial plan** I manage about 6 active production repos. I’m away from my desk a lot, so I wanted to manage stuff from my phone without opening a laptop. The original blueprint in my head was the classic setup everyone talks about: build a Telegram bot as the central brain. I text the bot, it pings Claude (Sonnet/Opus), Claude acts as the senior architect and delegates code writing down to a cheaper worker agent (Antigravity/Gemini), worker writes code, bot reports back.

**What actually happened** As I started messing around with Claude CLI on my main machine and looking at what was actually practical from a phone, o just have up and thought why not just use the Antigravity Android app itself as the UI?

It runs Gemini Flash natively, it connects directly to the repos on my main rig and it's fast and easy to use on my fold 7.. But Flash can be fucking obtuse sometimes. Every time Claude audits a change, there is something wrong.

So instead of Claude bossing Gemini around from the top, I made Gemini the front door, and gave it terminal access to run Claude Code CLI when things get serious.

**The setup right now:**

**The mobile front door:** I open the Antigravity app on my phone (running Gemini Flash). If I have a rough thought or want to look at a report while out and about, I just open the app and start a chat. 

**Calling the architect:** If the report or chat or whatever surfaces something worth building or fixing, I stay in the exact same chat thread and tell Flash: *"Hey, call Claude CLI to audit this idea and write a strict plan."* Flash runs a small wrapper script on my box. Claude Sonnet inspects the repo, catches the edge cases, gives an audit of whatever flash suggested and then if agrees something needs doing, it provides step-by-step implementation spec (which files to touch, interface changes, tests etc). That spec dumps straight back into my ongoing mobile chat.

**The worker pass:** I read Claude’s plan on my phone. Hit approve. Flash/Antigravity picks up the plan and does the work according to the olan. scaffolding files, inside an isolated git worktree on a bot/\* branch. It never touches main. Because it's following Claude's exact blueprint, it doesn't wander off or hallucinate weird patterns.

**The "fuck it, I'll do it myself" circuit breaker:** This is the most important part because I hate agent doom loops. Claude CLI audits the resulting git diff and test output. Flash gets **maximum two tries** to fix any errors or things Claude flags. If it fails the second pass, Claude stops reviewing, overrides the session, writes the fix itself, and commits it. Hardcoded in the bash script with exit codes so they can't sit there politely arguing about formatting until my credit balance dies.

**The Mac node fallback:** I got an M1 Mac sitting on the same Tailscale network running Ollama (local Llama). Both models can offload dumb background chores to it via terminal—summarizing huge diffs, formatting logs, checking git status. Free local compute so I'm not wasting cash.

**Why I kinda love it** It's all in one single chat thread on my phone. That's basically it. I do 90% of work on a laptop but sometimes because of life I only have my phone. So that means it's cool to have no app-switching, no copy-pasting between Claude web and an IDE, no squinting at microscopic fonts in a mobile SSH termina or reembering termux commands. l. I have an idea on the couch, Claude plans it, Flash writes it, Claude sanity-checks it, done.

**Where I'm second-guessing myself:**

I'm using three tiers of models (Flash + Sonnet + local Llama) just for 10% of my work. I mean antigravity has Claude models inside it which I could easily use (although my plan only allows for very limited usage of that over a week).

 I rely too much on Flash to know when something is "too complex" for it? (Right now I manually tell it to call Claude, but eventually I want rules like "if it touches auth or migrations, force Claude CLI").

 2 revisions is either too much or too little checks before takeover.

It's probably already been built, tested and improved as a standalone process from someone smarter than me and I could save time copying that.

In any case o figured it was cool and interesting to anyone else with either a pair of ar glasses or a very large phone screen.

document fetched 2026-11-03 00:41 · canonical https://reddit.com/r/ClaudeCode/comments/1xm9wse/connecting_claude_cli_to_antigravity_ide_for_a