Google Posted 'Reinstating Access.' Sixteen Days Later, the Auto-Ban Was Still Cycling. (The Thread Said Otherwise)

A discussion appeared on Google's Gemini CLI GitHub page and surfaced on Hacker News: "Addressing Antigravity Bans and Reinstating Access." 254 points, 216 comments. The topic was Antigravity, a third-party tool that let users route their Gemini subscription credentials through other AI clients, notably OpenClaw. Accounts had been getting suspended. Google posted a response. The thread underneath it shows a different picture than the title suggests.

What Antigravity actually was

Antigravity functioned as a proxy: pay Google for a Gemini subscription, hand your OAuth credentials to Antigravity, and it would let you use those credentials with OpenClaw or other clients outside Google's own ecosystem. The appeal was simple: one subscription, your choice of interface.

The reason bans started: Analemma_ laid out the mechanics plainly. OpenClaw, routed through Antigravity, was using "orders of magnitude more tokens than the IDEs." A flat-rate subscription is priced around what an average user consumes in a month. A bot running autonomously through a proxy is not the average user. Automation caught them.

This matters for the fairness question, which the thread doesn't answer cleanly. Some of those suspended were running OpenClaw bots through their subscription credentials. That's a straightforward ToS violation on the consumption side, and multiple commenters said so directly. Others had turned off tracking inside Antigravity and believed they were using the service within normal parameters. smashah put it plainly: "I turned off tracking on Antigravity. Do I deserve to get my account banned from a service I pay for now? Silly." The thread does not say how many people fell into which bucket. What it does say is that the automation that caught the high-consumption users also caught both kinds under the same flag.

We can't call these bans uniformly unfair without knowing which usage type triggered each account. What we can say is that the mechanism didn't distinguish.

Sixteen days, not a week

cube00 cut through the official framing. In response to the phrase "over the past week," they replied: "A week? Try at least 16 days," with a link to a separate Google forum thread on account restrictions. Then they described what the reinstatement process actually looked like: "The danger here is they'll ban you with no specific reason, fill out the form and you get an automatic unban and then something else automatically flags and you're banned the second time permanently. Support bot will then say 'you were warned, read the TOS' and you get to guess what you did wrong. You'll notice there are no appeals or reviews in this workflow."

The process is a loop, not reinstatement: automated suspension → generic form → automated unban → automated re-ban → automated "you were warned," with the second suspension permanent and no human in the chain. The official post title says "Reinstating Access." The process described in the thread is a cycle that delivers the same outcome on a second pass.

The pattern is similar to what we noted in the Wise account freeze: the company publishes a communication that describes resolution, the affected user's experience is an automated wall with no escalation path. That freeze ran two weeks. These bans ran at least sixteen days. In both cases the underlying account was paid and in standing.

Google's response versus Anthropic's: what the thread actually says

iepathos commented near the top: "Refreshing response from Google especially given the incompetence with which Anthropic has handled bans." The thread didn't add detail on the Anthropic side, so neither will we. Nothing in this discussion documents what happened there. The implication is that Antigravity users were being suspended across both platforms, and that Google's handling was considered the better of the two. On what grounds the comparison was made, the thread doesn't say.

The competing tool in comment seven

agentifysh wrote: "I wrote this as a work around to use my subscriptions for claude, chatgpt pro, grok from codex cli but seems like gemini is already broken and will require another approach after this," then linked to their own GitHub project doing essentially the same thing as Antigravity, built for different subscription credentials.

We noted this move in the one-person agency thread and named it the stealth-ad in the empathy comment. Here it's less disguised: a thread about subscription bans becomes a discovery surface for the next subscription workaround tool. The timing is useful for the commenter. Whether they disclosed a financial interest in their own tool, the thread does not say.

The structural problem the thread is pointing at

xrd took the widest angle: "if AI is going to replace all employees, what guarantees do you have as the employer that your employees will do your bidding, and not the bidding of enterprises with a shifting moral landscape? Once we have tooling wrapped around specific agents, it'll be hard to rehire. What will we do then when our 'employees' are furloughed?"

The framing is speculative, but the immediate problem is real. If your workflow depends on a subscription account at a platform that can suspend it automatically (with no warning, no stated reason, and no human appeal), you don't own that capability. You are renting it, and the lease has clauses enforced by automation you cannot reach.

A commenter in an earlier thread we covered lost a TikTok account: 1,750 followers, a ban they couldn't explain, no appeal window for thirty days. The income that survived came from channels they owned outright: an email list and a blog. Their framing: "they can't remove my email list, my blog and all others as assets I own." The platform risk is the same shape applied to professional tools instead of social audiences. The rented channel disappears; the owned one keeps running.

The Antigravity thread adds a detail the TikTok example didn't have: you can be reinstated, pass the automated check, and be suspended again before the week is out. That makes the reinstatement not a resolution but a pause.

What the thread does not tell you

The discussion is largely philosophical and technical: debates about whether subscriptions should allow any client, whether locking to first-party tools is anticompetitive behavior, what the ToS actually permits. There are no concrete reports of lost client work, missed deadlines, or revenue numbers tied to the downtime. We are noting the absence, not explaining it. The thread may not include those voices, or those voices may not have been in this discussion. Either way, the data does not support a claim about how much work actually stopped.

What the data does support: at least sixteen days of access suspension for some users, an automated cycle with no effective appeal path, and a ban mechanism that caught high-consumption ToS violators and users who believed their usage was within bounds under the same automation with no stated distinction.

The tool, the catalog, and what the community is filing bugs about

Gemini CLI (the official Google client at the center of this ban wave) is in our toolbox alongside the broader AI coding agent ecosystem. That page tracks star history, top issues sorted by community votes, and neighboring projects in the same category. If you're choosing between clients for any AI subscription, the issues list shows what the community is actually filing bugs about, which is a different data source than the feature comparison on the landing page.

The account-risk-audit prompt in the panel does the dependency mapping before something like this reaches your account: every tool labeled by what stops if access disappears tonight, single points of failure identified, substitution readiness rated. The tool-portability-check puts a number on how locked in you are: hours to migrate, not a vague "it's complicated." The ban-appeal-drafter turns the emotional first draft into something with enough facts that a human reviewer can act on it, assuming one ever reads it. And the downtime-continuity-plan covers the first 48 hours when it does happen: what to triage, what to tell clients, what to reach for.

The subscription-as-lease model is not going to change. The labs price their flat-rate plans on average consumption and enforce the terms when something uses far more. That logic is coherent from their side. The user's question is a different one: which of my workflows stops if the account goes down tonight, and which of those have I actually thought about?