Launch HN: Maritime (YC F26) – A managed cloud for production AI agents
Hi HN, we're Ary and Maria, co-founders of Maritime (https://maritime.sh). We're building an agent-native cloud that handles everything from custom environments to lifecycle management for AI agents.
Demo: https://www.youtube.com/watch?v=lCWMdBQ96rY
There’s been an explosion of sandbox companies coming out of SF, but sandboxes solve maybe 10% of the infrastructure problem for AI agents. The approach that most sandboxing companies take is:
You buy Hetzner/OVH Cloud bare metal servers, use Firecracker to slice the servers up into VMs, sell those VMs with low margins.
It’s a bit of an oversimplification, but the point is it’s a narrow solution to a fairly difficult problem.
We talked to a couple hundred companies that are building AI agents. From personal assistants and coding agents to specialized healthcare agents, and the main pattern we saw was that even after choosing a sandbox provider, you still have to build additional infrastructure just to get things working. These were the primary points of friction:
1. Agent lifecycle management.
There’s a lot that you have to build in order to properly give your agents VMs: things like sleep/wake, persistent storage, triggers, scheduling, per-user routing and spend caps, health monitoring, failure recovery, and deploying updates across thousands of VMs without downtime. There are so many of these small sandbox management tasks that companies have to build and maintain themselves, and they add up.
2. Custom environments.
Production agents have varying needs. Some agents need Windows applications. Others need Android, browsers with persistent sessions, custom operating system images, preinstalled desktop software and applications, or specific dependencies. Some agents need access to private networks to interact with customer-specific things on prem, or residential IPs, or the ability to restrict network access completely.
3. Cost optimization.
If you’re selling a personal assistant for $20/month, a sandbox that costs $0.04/hour comes out to roughly $29/month if you keep it running continuously, not including LLM inference, storage, etc. There is a way to combat this though, which is to optimize sandbox usage with auto-snapshotting, making things event-driven, mounting S3 buckets to agent environments via FUSE for cheap storage, cold-tiering snapshots, etc. All roads lead to… building more infrastructure.
So what did we build differently?
We give each agent its own isolated, persistent microVM and a configurable environment. On top of the compute, we built a lot of composable services: provisioning, persistent state, sleep/wake orchestration, triggers, schedule creation, monitoring, network configuration, updates, rollbacks, recovery, per-agent spend tracking, login handoffs, all the things you’d need for production agents.
We also spent a huge amount of time optimizing cost. A lot of thought was put into things such as scale-to-zero, copy-on-write over shared base images, compressed memory snapshots, hot/warm/cold snapshot and disk tiering, and other optimizations while balancing these with performance metrics like latency.
As a result, Maritime costs $1 for a 2gb RAM agent per month (unlimited usage for that $1), with additional resources available as needed.
Maria was a visiting lecturer at MIT and did research at Project NANDA on agent infrastructure. Ary did research on agent coordination protocols at the same lab. We started Maritime together this March because of all the issues we faced with trying to deploy agents to the cloud during our research.
We want your feedback.
If you're building an agent product, what did you have to build on top of sandboxes to make them work? We want to know about all the workarounds and weird things you’ve had to build.
Are there things you’ve tried to build with VMs but couldn’t? Does Maritime fit your needs? Is there a feature we’re missing that would make us more enticing?
Happy to answer any questions in the comments.
Congrats on the launch. One question the "sandboxes are 10% of the problem" framing raises for me: with each agent in its own VM, what does the agent get for outbound network and for credentials?
The two failure modes I would want a managed platform to handle by default are (1) an injected instruction in a fetched page or a customer email telling the agent to send data somewhere, and (2) a long-lived provider key sitting inside the VM where the agent, or anything it runs, can read it. Two patterns that seem to hold up: an egress allowlist enforced outside the VM rather than in the agent's own config, and giving the VM a scoped, revocable token that a proxy exchanges for the real key, so a leaked token has a spend cap and a kill switch.
Do you do either of those today, or is that left to the customer's harness?
Congrats on launching! I think the lifecycle management and cost-optimization layers you're building on top of microVMs address some painful problems for teams deploying agents in production.
I work at a neocloud company, and from my experience, getting VMs up and running with Firecracker/KVM on bare-metal infrastructure is relatively straightforward and requires little operational overhead initially.
The real challenge is what comes afterward: continuously monitoring, patching, and responding to newly discovered CVEs. Falling behind on security updates can have serious consequences, and ultimately, it's your customers who bear the risk.
This is one reason I'd be cautious about choosing a smaller provider based solely on price. At the infrastructure level, many providers are essentially selling the same isolated VMs. The real differentiators should be security, reliability, and operational maturity, particularly with AI accelerating vulnerability discovery and expanding the attack surface. Saving a few dollars on compute might not be worth the trade-offs.
There's also the question of geography and latency. If your agents run in a big tech cloud, offloading sandbox execution to infrastructure on another continent isn't necessarily a great choice. The added network latency can significantly impact agent responsiveness, especially when agents need frequent interactions with external services or real-time applications.
Thank you! And yes I agree that getting VMs up is quite trivial, patching hosts and guest kernels after CVEs is part of what we mean by "managed" - we don't want our customers having to continuously maintain things.
As for latency, we've noticed that since most agent runs take >10s, a couple extra ms of latency isn't a huge problem for a lot of companies. That said, we're planning on getting more edge hosts, and BYOC.
Did you guys … generate this slop for engagement?
Nope
They didn't, they asked their friend to do it for them.
The account is 3 months old and has only posted slop like this on launch threads.
Forgive my naivety, but if it's built for agents then why should humans need to login, view a dashboard, care about it?
Thinking about product market fit, buyer personas, and right audiences here.
Would there be a future where maybe the agent can elevate to the human what the human needs to know when it wants to or requires it from what it has access and these platform-like UX constructs are more like a "just in time" construct similar to Claude Artifacts?
I expect the next cloud for AI to be fully headless, though maybe I'm wrong
We are headless! We spent a lot of time fleshing out our CLI so our customers' coding agents can deal with things themselves. That said, a lot of customers want some sort of visual basis to see what's going on, especially if they have a lot of users.
I might try your platform but I am kind of confused at the pricing.
Let me explain my current setup. We have something like muse/instinct/dots equivalent built for our customers(GTM domain), so each of them gets an agent with a VM (sandbox - currently using daytona). The sessions and the context (skills, instructions, hooks etc.) are synced to R2. Daytona(2GB ram, 3GB disk, 1vcpu) is working out good for us, cold starts are <500ms, whereas warm ones (non-archived sanboxes) are much lesser. Current problems we are facing is around agent swarms or sub-agents or multiple sessions from same customer (teams trying to interact) - we needed to manage the daytona sandbox sizing on this.
If I were to use maritime, what'd be different for me? Would I be paying $1/mo per customer agent? I can just move my harness to maritime and stop managing sandboxes, R2 sync etc.?
Another company I wanted to checkout is Omnara(another YC company), but Maritime felt closer to my task.
Thanks for explaining! I’d start with one persistent agent box per customer.
It’s $1/month for 1 vCPU, 2 GB RAM and 5 GB disk. Plans start at $20/month, including $20 of usage and 5 machines awake simultaneously. Model costs are separate.
You can bring your harness as a Docker image with a small HTTP wrapper. State saved under /data persists, so R2 isn’t needed just to restore that workspace. Keep it for backups or sharing state between workers.
Subagents sharing a box share its resources. Separate worker sandboxes add usage charges. We manage starting, sleeping and waking machines, but your harness still handles session coordination and worker sizing. Cold starts are <500ms as well
Permission rules are only half of it. On my own homelab setup, a config linter found 6 MCP server tokens stored in plaintext in ~/.claude.json, even though the permissions were tight. Allowing a whole MCP server also allows its later write tools, and unpinned npx/uvx packages pull the newest version on every start. Wildcards are another trap: Bash(python3:) looks narrow but is effectively Bash(). For agents that change networks or cloud resources, I keep approval gates outside the agent and bind each approval to a specific plan.
The 429 caps the model bill. The expensive mistake is usually the tool call, though: a refund, a bulk email, a Terraform apply. Those leave the VM and go to the customer's own APIs, and a model budget never sees them. Is outbound budgeting per agent on your roadmap, or do you see it as the harness's job for good?
The point about sandboxes being only a small part of the problem makes sense. Lifecycle seems like where most of the custom code ends up. Two questions: how do spend caps work when an agent gets stuck in a loop mid-task, do you stop it hard or let it reach a checkpoint first? And when a VM sleeps and wakes, what state is kept beyond the disk, for example open connections or in-progress tool calls?
Spend caps are enforced by us returning a 429 when a model provider is called, you can elect to have some sort of loop detection but that comes down to harness engineering - we don't touch what runs inside of the VM.
In-progress tool calls are kept upon sleep/wake, however, open connections aren't (since typically the other side usually times out either way)
Nice that in-progress tool calls survive sleep/wake. The case I'd worry about is the unplanned one. If a host dies or you roll an update, does the box come back from its last memory snapshot or cold from /data? Coming back from a snapshot means the agent has forgotten whatever it did since, so it can redo a push, a message or a charge that already happened. Since you leave the harness alone, do you at least signal to the VM that it was restored, so the harness can check before repeating a side effect?
More slop from a new account. This launch thread is a joke.
More slop from a new account in this thread.
Really, you think I talk like an AI? Interesting.
On the spend cap question, the thing that bit me was that "let it reach a checkpoint" only works if your checkpoints are cheap to reach. An agent stuck in a loop can burn a lot between checkpoints, so I'd want to know whether your cap is enforced per-call (refuse the next request once you cross the line) or per-task (reconcile after the fact). The first is the only one that actually prevents overspend; the second just tells you about it.
Worth being explicit in your docs about which one it is, because people will assume the hard version and then get surprised by a bill from a runaway loop.
For the enforcement layer specifically, I built CostWindow for this — a drop-in proxy that refuses the call once a process hits its budget, before the provider is charged. Might be redundant with what you're doing internally, but it's the pattern I'd reach for.
It’s been cool to see the evolution from vms to serverless agents.
I think it’s the right kind of innovation sort of like ec2, serverless functions, and etc in the age of ai
Thank you! I think this is where things are headed too
How do you compare yourself with https://www.agent37.com/
Also, digitalocean offers this now.
Better landing page animation, of course.
Agent37 is great if you want plain, cheap sandboxes. But you still have to build and manage a lot of the infra to sell agents to customers, we manage the entirety so that you don't think about cloud runtime at all and can focus on other things. We also provide you with ongoing support, an infra partner if you may.
DigitalOcean is focusing on coding agent sessions in the cloud, rather than managing the things necessary to sell long living agents to customers.
There are a lot of alternatives in the market for this.
Most are focused on providing commoditized compute, there's a lot more nuance when building an agent company.
I dont think there is [1]
It is exactly the same offering.
[1] https://www.digitalocean.com/products/managed-agents
Like what?
Here a list of harness, sandboxes, paas doing the same [1]
[1] https://debarshibasak.github.io/awesome-paas/
A cloud box doesn't really make much sense to me (personal opinion): monthly VM cost + LLM cost is more than a canned personal assistant; or host a agents on a mac mini at home and only pays for the AI subscriptions, which is what I am doing now.
cool product. right now I just built my own version of this on AWS (which is likely far more expensive, but we have startup credits). I'll definitely check this out one day when the credits end/expire. I will say that startup credits would really help!
side note, this design aesthetic is so cool but now literally every SF AI startup uses it -- I'd almost prefer craigslist style where the information is to the point and direct. i.e how much money/time will this save me and is it secure. the two questions i only care to answer here. i think the aesthetic is meaningless at this point bc of how easy it is to vibe this template out
Fair! What did you build on AWS? What was the most time consuming part?
its basically a 24/7 box with claude code on it. I think the most time consuming part was not the box setup, but rather getting my local skills/workflows (that are fairly scattered) into something rebuildable/reusable. And as I work on my box, i realize there are missing dev pieces that I had on my local machine so I always have to keep remembering to update the docker image with the base level CLIs/MCPs/skills that i need. Wasn't too bad, but definitely annoying
Do you keeps a per-run record of tool calls that a team can query, so the runs with looped calls can be found before the cap trips?
what happens when agent reads mails with hidden instruction and how do you control agent's actions
How are you going to pay back the 500k if you’re charging $1/month and competing with a thousand other startups that are all racing towards the bottom?
We're not racing to the bottom because sandboxes are a small part of our offering. Every company is going to be running thousands of agents sooner or later, and they'll all need custom environments. More agents = more $
Cool :) if I used Maritime to power my agent product, would my end users be able to sign in with their existing AI subs? Or would they have to enter their API key.
Up to you, we have support for BYOK, AI subs, or pass-through where you can top up our LLM gateway and we take care of spending caps per user, etc. Whatever fits your use case best. What are you building?
there is no such thing as bringing AI subs, it's against TOS on Claude and OpenAI to use your subscription in third party tools. You have to use API key usage which is extra on top of your subscription
Congratulations on the launch. Very interesting idea. I read your description, and it's striking: it's as if you described our infrastructure at Playcode (https://playcode.io/cloud). The similarity is just colossal, it's what we built for ourselves and for our clients.
Its like sandboxes are a popular space or something... who would've thought. its surprising people have similar infrastructure to playcode I truly can't believe it
what kind of cost optimisation support you guys plan to give? Will it be similar to what say AWS offers today at EC2 and other resource levels except you guys on token?
Is this similar or using CNCF's agent-substrate open source platform?
I think the primary distinction here is we let each agent get its own isolated VM. Agent substrate lets you run many agents on shared workers. Plus we run the whole thing for you :)
i've been looking into maritime, they're heavily customizable plus it has a really good sdk and interface if you want to integrate it with you existing platform.
Thanks!
What do you do we don't get with kagent?
kagent gives you the framework, and you run it on your own k8s. We’re the managed service that runs all the infra for you
how to provide tools to the agent ? the agent in my product manipulates application state/updates databases etc
You'd define the tools in the agent harness, and you can set up env variables through us. You can also connect to on prem databases using things like our connect module.
I just run Claude Code sessions in tmux on my Mac and talk to them over Telegram. I don't really see what a managed cloud adds for most people.
We sell primarily to startups that run agents, rather than directly to individuals. E.g. healthcare agent companies, personal assistants, products with embedded agents, etc.
sounds like the dropbox comment
;)
Same setup here (Claude Code in tmux, Telegram for messages), and for one person it works. Three things I ended up building by hand, though: handoff notes before clearing context, because a long session loses the thread (I've written about 97 this month); a small VPS for the parts that must run when my laptop sleeps (cron jobs, alerts, a dashboard sync); and a pre-tool-use hook that blocks certain commands, since the agent otherwise has all of my permissions. Next I'm trying herdrchat, an open-source iOS client that reaches the terminal sessions over SSH with no relay in between, to replace Telegram for phone access. None of that is hard for me alone. For a company running one agent per customer, it is the lifecycle work their post describes, multiplied by the number of customers.
the handoff notes before clearing context are the part i would not skip. i keep a short handoff.md with what's in progress, what got tried and dropped, and the next step, otherwise the next session just rebuilds the same context from a noisy transcript
Same here. The laptop-sleeps problem is what pushed me off "just tmux". I ended up with agents on a desktop plus a cheap box, and my laptop is mostly a remote control. The annoying parts were keeping project files in sync and reaching sessions without opening ports or relaying through someone's cloud.
I built an open-source tool for this: Beep (https://github.com/flashmind-labs/beep). It runs Claude Code / Codex / OpenCode / Pi across your own machines over encrypted p2p (iroh/QUIC). There's no account and no middle server, it does Syncthing-compatible project sync, and it uses your existing subscriptions. If you do want managed infra, Maritime makes sense. But for one person with a couple of machines, you may not need a cloud at all.
That was my setup too, with my own harness on a Mac mini at home, which took care of the sleep problem, and Telegram on my phone. I moved off Telegram app because I wanted full control of the harness and agents that could talk to each other.
Is it bring your own FS?
No. We mount an individual /data volume to every VM that persists, but if you have data that's in S3 or R2 you can hook those up as well.
You could stand to hire a copywriter for the product pages. It reeks of being AI generated, and these days that causes me to kinda sigh and not really pay attention. Which might be totally unfair to your product! But it's how I feel.
Right now I'm using Grains.run -- it has toolbox, x402, custom tools, fs, memory, and more
plus it basically plugs into everything out of the box (ides, sdks, apps, etc), I haven't even broken the free tier for all my loads thrown at it.
It's reliable enough we even use it in production systems.
might give it a try, great work
thanks! let us know what you think