Maritime wants to be your agent's landlord, and the rent is $1 a month.
ngl the cheap part is the least interesting thing about it.
Maritime, a YC F26 startup from co-founders Ary and Maria, did a Launch HN on November 18 for what they call an agent-native cloud. Each agent gets its own isolated, persistent microVM, plus the layer around it that nobody wants to build: sleep/wake, triggers, scheduling, monitoring, updates, rollbacks, recovery, per-agent spend tracking and login handoffs. The thread had 55 points and 49 comments when we pulled it.
hot take: their diagnosis is correct and their moat is not.
sandboxes are the easy 10%, and they're right about that
The pitch opens with a spicy line: sandboxes "solve maybe 10% of the infrastructure problem for AI agents." The founders say they talked to a couple hundred companies building agents and kept hitting the same three walls. Lifecycle management. Custom environments, like Windows apps, Android or a customer's private network. and cost.
The cost math is the best part of the post tbh. Sell a personal assistant for $20/month, keep it on a $0.04/hour sandbox around the clock, and that's roughly $29/month before you've paid for a single token. You're underwater on compute alone.
Maritime's answer is to let agents sleep. Their homepage says one AMD EPYC 7502P server (64 threads, 503 GB) held 1,107 agents with 29 awake at once, counted 4 Sep 2026, and a paper excerpt on the same page says a typical agent of theirs is awake for eleven minutes a day. That's the trick behind $1 a month for a 2 GB agent box: scale-to-zero, copy-on-write base images, compressed memory snapshots and hot/warm/cold tiering. Desktops, phones and throwaway sandboxes are billed per hour awake.
my read: this is the right abstraction. An agent is a long-lived process with a disk and a schedule, and teams keep hand-rolling the same sleep/wake glue over and over.
The sharpest pushback came from a commenter at a neocloud: getting Firecracker VMs up is easy, the real job is patching CVEs forever, so picking a small provider on price alone is risky. The founders replied that patching host and guest kernels is part of what "managed" means for them. They also said most agent runs take over 10 seconds so a few extra milliseconds of latency rarely matter, and that more edge hosts and BYOC are planned.
Then came the "isn't this already a thing" pile-on, with commenters naming Agent37, DigitalOcean's Managed Agents, CNCF's agent-substrate, kagent and Playcode. DigitalOcean's page (public preview) lists a Firecracker microVM per session and auto-pause on idle, which is kinda the same shape. Maritime's replies: DigitalOcean is focused on coding agent sessions, and agent-substrate runs many agents on shared workers while Maritime gives each one its own VM.
Someone running Claude Code in tmux on a Mac said they don't see the point. Another commenter instantly called that "the dropbox comment." iconic.
imo the moat question is the real one. $1/month is a price, not a position, and a big cloud shipping a similar runtime is the scariest thing in that thread. Maritime's bet is that the managed lifecycle and the support are the product, not the VM.
the sleep/wake bug that should keep you up at night
The most useful exchange was about state. Maritime said in-progress tool calls survive a sleep/wake but open connections don't, since the other side usually times out anyway.
A follow-up asked the scary version. If a host dies or an update rolls out and the box comes back from an older memory snapshot, the agent can redo a push, a message or a charge that already happened. Does the VM at least get told it was restored? That one was still unanswered when we pulled the thread.
Say your assistant sends an invoice reminder, then the box restores to a snapshot from five minutes earlier. It sends it again. fun.
Writes need idempotency keys or a check-before-act step on any cloud tbh For what it's worth, MewCP's Hermes setup docs note that Hermes won't auto-retry a tool call that might change something if its MCP connection expired while it sat idle, and returns outcome_uncertain instead. Right instinct. And when calls go through MewCP, every one lands in a log with the tool, time, status and latency, so you can see whether that send went out before you try it twice.
Someone asked how to give the agent tools that change app state. Maritime's answer: you define tools in your harness, set env variables through Maritime, and use their connect module for on-prem databases. So the compute is managed and the app access is still on you.
That gap gets real fast with a use case from Maritime's own homepage, one assistant per customer on WhatsApp, Slack or email. Say each customer's assistant needs that customer's Gmail and Google Calendar. Now you're storing and refreshing hundreds of OAuth tokens and making sure customer A's agent never touches customer B's inbox. Env vars are a rough home for that.
This is where I'd put MewCP next to it. Maritime runs harnesses like OpenClaw, Hermes, the Claude Agent SDK and OpenAI Codex, and MewCP has setup guides for OpenClaw and Hermes, so the agent in each box holds one MewCP key instead of a pile of app tokens. Each customer connects their own accounts, your backend passes their user id, and every call runs on that person's accounts and nobody else's.
stuff we still don't know
How a restored box tells the harness it was restored. Unanswered.
Whether "unlimited usage for that $1" holds up for agents awake way more than eleven minutes a day. The whole cost model is built on most agents being asleep.
Maritime says end users can bring their own key, use AI subscriptions or top up its LLM gateway. One commenter claimed using a Claude or OpenAI subscription in third-party tools breaks their terms. We didn't verify either side, so read the terms yourself.
where this came from