Grass: Run Claude Code on a Dedicated VM, Monitor from Phone

Grass launched on Product Hunt with a pitch that fits in a sentence: run Claude Code on a persistent VM and check on it from your phone. That’s it. That’s the product.

It’s a narrow premise, and it’s almost certainly the right one to start with.

Anyone who’s used Claude Code for extended sessions knows the failure mode well. You kick off a refactor that’s going to take a while, step away, come back to find a timeout error staring at you. Or you’re not at your desk at all and you want to know if the agent is still running, still making progress, still on the right track. Checking in from your phone currently means SSH tunnels, terminal apps, and a lot of squinting. Nobody designed that experience on purpose. It’s just what you’re left with when mobile was an afterthought.

Grass is betting that enough developers share that frustration to make a product out of fixing it.

The infrastructure side

Instead of running your coding session on local compute, Grass provisions a VM for you. The VM runs 24/7. It’s pre-configured. You don’t set it up. You point Claude Code at it as you normally would, except the compute isn’t your laptop, and you can “monitor from your phone” mid-session, steer it, push changes, all without touching your machine.

The VM layer runs on Daytona, an open-source development environment manager. Daytona handles the infrastructure orchestration, which means Grass isn’t building bare-metal VM management from scratch. For an alpha product, that’s a smart call. You get fast environment spin-up, clean developer surfaces, and you don’t spend six months reinventing wheel infrastructure before you’ve validated that anyone wants the thing on top.

The BYOK model, bring your own key, is worth paying attention to. Your Anthropic API key stays on your side. Grass doesn’t touch it. That’s not a trust-building gesture so much as it’s the obvious correct answer to a question developers will ask immediately. Nobody wants to route API credentials through a service they found yesterday. Keeping the key out of Grass’s hands removes that concern before it can become a blocker.

Every new account gets 10 free hours. No credit card required. That’s a low enough barrier that there’s almost no reason not to just try it.

Agent support

The agent-agnostic framing is real but currently limited. Claude Code works today. OpenCode works today. More agents are listed as coming. That’s a reasonable place to be in 2025, when the market for AI coding agents is fragmented enough that committing hard to one runner would be a mistake you’d regret inside of a year.

OpenCode is worth flagging specifically. It’s an open-source terminal UI for running Claude that’s been getting real traction as an alternative to Anthropic’s official client. Including it signals that Grass is paying attention to what’s actually gaining adoption in the developer community rather than just supporting the obvious name-brand option.

The mobile angle

Most developer tooling treats mobile as a concession. You get a web interface that technically renders on a phone, it works if you squint, and that’s considered coverage. Grass is positioning phone access as a first-class feature. The difference matters.

The GitHub repo is an Expo project, which means the team is building with React Native for native mobile. An App Store release is listed as coming soon, which puts the current experience on a web-based interface. But the direction is clear.

I asked the team about what they think they’re actually building here, and the framing they gave me was specific. The goal, they said, is to “make the right environment for the agent to live in and let the human interact with it asynchronously.” That’s a different mental model than most coding tools, which still assume you’re sitting at a keyboard, watching the output scroll. Async interaction with a running agent, from wherever you happen to be, is a real shift in how you’d actually work with these systems over a full day.

Who’s building this

Grass describes itself as a “small team that actually wants to talk to users.” That’s the kind of thing every early-stage team says, but the Product Hunt comments suggest they’re following through on it. They’re responding to questions, adjusting based on feedback, and treating the launch as the beginning of a conversation rather than a press moment.

They’re also direct about what the product is for. The stated mission is to “make AI coding better.” That’s not a positioning statement with a lot of give in it. They’re not claiming to reinvent software development or replace the programmer. They’re trying to fix the specific, annoying gap between AI coding agents and the reality of how developers actually move through a day.

The 7-day question

What changes if you use this for 7 days? That’s usually the right frame for evaluating developer tools. Day one is setup friction. Day 30 is habit. The interesting signal comes in the first week, when the novelty has worn off but the tool hasn’t become invisible yet.

With Grass, the 7-day test is really a test of the async model. Does offloading compute to a persistent VM actually change how you work, or does it just move the problem around? The timeout issue is real, but it’s one symptom of a broader pattern: AI coding agents were built for single-session, local use, and the workflows around them haven’t caught up to what people are actually doing with them.

If Grass’s VM layer genuinely holds sessions stable across hours, and if the mobile interface is good enough to steer an agent without wanting to throw your phone, then the core premise holds. If the VM has its own failure modes, or if the mobile experience is web-responsive-but-not-really, then you’ve traded one set of annoyances for another.

Context

Anthropic’s Claude Code has been one of the more widely adopted AI coding tools since it launched. The developer audience for it is sophisticated. They know what they want, they notice what’s missing, and they build workarounds when the official experience falls short. Grass is essentially a formalized workaround that bets the workaround is valuable enough to pay for.

The 24/7 persistent compute angle is already table stakes in some adjacent categories. Cloud development environments have been around for years. What’s new is the combination: persistent cloud compute, AI coding agent running on it, and async mobile access as the primary interaction model. None of those three pieces is individually novel. The combination hasn’t been packaged this way before.

Whether the market is big enough to sustain a standalone product around this specific gap, or whether it gets absorbed into a larger platform, is not a question Grass needs to answer right now. They need to answer whether the thing works and whether enough developers find it indispensable in the first 30 days of use.

The 10 free hours are enough to find out. The React Native app, when it ships, will be the real test of whether the mobile-first framing is genuine or cosmetic.

The team told Product Hunt they want to hear from anyone who tries it. That’s a low bar to clear. But clearing low bars consistently is how products get better fast enough to matter.