What Is Maritime? The Beginner's Guide to Giving Every AI Agent Its Own Cloud Computer
CodexFuse Team · 2026-11-05T22:03:23.457Z · 20 min read min read
What Is Maritime? The Beginner's Guide to Giving Every AI Agent Its Own Cloud Computer
What Is Maritime?
What is Maritime AI agent hosting? Maritime (maritime.sh) is a cloud platform that gives every AI agent its own persistent computer, an isolated Firecracker microVM with its own disk, memory, files, and logins that survive between tasks. The machine sleeps when the agent is idle and wakes in about a second when a message arrives, and Maritime charges $1 per agent per month for the box.
Maritime (maritime.sh) is a cloud company for AI agents, not a shipping company. The bare word Maritime mostly returns shipping news in search results, so this article always uses the qualified name, and it covers the YC Fall 2026 startup founded by Maria Gorskikh and Aryaman Shrivastava.
The idea is simple. A personal assistant, a coding agent, or a whole fleet of agents each needs a place to live, somewhere its files and logins wait between jobs. Maritime gives each one that place, instead of renting a full server that sits idle or a serverless function that forgets everything between calls.
Agents keep API keys and logins on their own disk, so the same care you give a password applies here. A SHA-256 hash generator is the quick way to check a file fingerprint before you trust it, and it is the kind of small verification habit that matters when an agent holds your credentials. If you have ever rented a server, the contrast is immediate: VPS hosting keeps one machine awake around the clock, while Maritime parks the agent machine until a message arrives.
| Maritime at a glance | Detail |
|---|---|
| What it is | Cloud computers for AI agents, one persistent Firecracker microVM per agent |
| Company | Maritime (maritime.sh), YC Fall 2026, founded by Maria Gorskikh and Aryaman Shrivastava |
| Headline price | $1 per agent per month for an agent box, three agents free |
| Wake time | About a second, per the vendor, after sleeping while idle |
| Computer types | Agent boxes, sandboxes, swarms, Linux desktops, Windows desktops, Android phones |
| Control paths | Run code directly, MCP, dashboard chat, CLI, REST API, Telegram, cron and webhook triggers |
| Scale claim | Maritime says it hosts thousands of agents for individuals, students, small teams and startups |
Why AI Agents Need Their Own Computer
A persistent agent is a process with a disk, a set of credentials, and a schedule of its own. That is how the founders describe it in their own research paper, and it is the whole problem: unlike a request handler, an agent has state that outlives any single interaction, which rules out the stateless scaling model of serverless platforms.
Renting an always-on server for every agent is far too expensive for anyone but a large company. Maritime's own blog puts it plainly, and the numbers back it up: a typical agent is awake for eleven minutes a day, according to the vendor's snapshot-restore paper from 3 September 2026.
Eleven minutes out of 1,440 minutes in a day is about 0.76 percent of the day awake, and 99.24 percent asleep. Over a 30-day month that is 330 minutes, or 5.5 hours, of actual work against 714.5 hours of sitting idle, which is why paying for a machine that never sleeps wastes nearly the whole bill.
Serverless looks cheap until the agent needs memory. A function forgets everything between calls, so every login, file, and half-finished task has to be rebuilt or reloaded from somewhere else. Containers fix the memory problem but bring back the DevOps work: someone has to run the host, watch the disk, and restart the crashed process.
Maritime's answer is to keep the machine and park it. The agent's microVM sleeps to disk when idle, so a parked agent costs only its storage, and it restores on the next inbound message. That trade is the entire product, and everything else Maritime sells is a variation on which kind of computer the agent wakes up on.
How the Persistent MicroVM Works
A Maritime persistent agent computer is a Firecracker microVM, a tiny virtual machine with its own disk, memory, files, and logins. Everything under the agent's data directory survives restarts, redeploys, and sleep-wake cycles, so the agent picks up exactly where it stopped.
The restore path has three steps, described in the vendor's own paper. The memory snapshot is mapped rather than read, so the first instructions execute before the file is resident. The block device is attached as a thin clone of the parked volume, and the network tap is re-created with the address the guest had when it slept.
The result is a machine that behaves like it was never off. The agent keeps its files, its memory, and its logins between tasks, and the next message reaches the same agent on the same machine, not a fresh copy that has to re-learn the job.
No independent benchmark of the wake time exists yet, so treat the sub-second wake as the vendor's own claim. The blog says the microVM wakes in under a second, the company page says about a second, and no outside tester has published a measurement either way as of 6 November 2026.
The Six Computers Maritime Offers
Maritime sells six kinds of computer on one API and one bill, and the right choice depends on what the agent does. Maritime cloud computers AI agent hosting is a menu, not one product: each AI agent cloud computer on it is a different shape, and the persistent box is only the first item.
| Computer | What it is | Best for |
|---|---|---|
| Agent Boxes | The persistent home: one agent per box, asleep until messaged | Personal assistants, one agent per customer |
| Sandboxes | Throwaway code execution that ends itself after the task | Coding agents running generated code safely |
| Swarms | One command fanned out to hundreds of sandboxes at once | Evals and reinforcement learning runs |
| Linux Desktops | A computer with a screen for apps with no API | Computer-use agents driving a browser or desktop app |
| Windows Desktops | A Windows desktop in the cloud, licence included | The Windows app your customer runs |
| Android Phones | A phone in the cloud, with a person able to take over by link | Apps that only exist on mobile |
Agent boxes are the flagship. Maritime offers templates for Hermes, OpenClaw, Claude, and OpenAI agents, and one box serves one customer, sleeping when idle and waking on the next message.
Sandboxes are the opposite shape. A coding agent runs generated code in a throwaway sandbox and keeps the repo in a workspace that is still there next week, while the sandbox itself ends after the task. Code written for the E2B SDK runs unchanged on Maritime sandboxes, according to the vendor.
Swarms scale the sandbox idea sideways. One job runs across 1,000 sandboxes at once for evals or training runs, and every exit code and output comes back for collection. At the 2 vCPU sandbox rate of $0.022 an hour, 1,000 sandboxes running for one hour cost $22, which is the kind of arithmetic that decides whether a swarm is cheap or ruinous.
Desktops and phones cover the agents that need a screen. A Linux desktop drives a browser behind a login, a Windows desktop runs the app with no API, and an Android phone runs the app that only exists on mobile. A person can take over the phone by link, which is the escape hatch when the agent gets stuck.
How to Give Your Agent a Computer in 5 Steps
Maritime's own guides describe a short path from nothing to a running agent, and the steps are the same whichever computer you pick. The work is choosing the computer and the framework, not learning a new deployment system.
- Pick the computer type: an agent box for a persistent assistant, a sandbox for throwaway code, or a desktop or phone for screen work.
- Choose a template or bring your own framework: bind the port Maritime gives you, answer a health check, and answer chat messages, and the platform handles the rest.
- Add your model key: the Claude Code template stores the Anthropic key encrypted and injects it only into the agent's own VM, so model usage bills to your own Anthropic account.
- Connect a channel: dashboard chat, the Maritime CLI, the REST API, Telegram, or cron and webhook triggers that wake the agent on a schedule or an event.
- Let it sleep: the box parks itself when idle, and your next message wakes the same agent with its files and logins intact.
The framework contract is deliberately small. OpenClaw, ZeroClaw, Hermes, and Claude Code all fit it, and the vendor's guides walk through each one with exact commands. If your agent already runs somewhere else, the move is a redeploy, not a rewrite.
Control also works the other way around. You can use MCP to let an existing agent control a persistent Maritime computer, so the agent you already have gains a machine without moving house. The Python SDK is on PyPI as maritime version 0.7.1, and a Files API with eight endpoints lets you browse, upload, download, and edit an agent's disk over HTTP, even while it sleeps.
Maritime Pricing: What $1 per Agent per Month Buys
Maritime pricing starts at $1 per agent per month for an agent box, with three agents free, according to the vendor's blog and company page. At that rate one agent costs $12 for a full year, and the three free agents are worth $36 a year at list price.
| Item | Price | Notes |
|---|---|---|
| Agent box | $1 per agent per month | Three agents free; persistent, sleeps when idle |
| Sandbox, 1 vCPU and 2 GB | $0.011 per hour | Billed per second |
| Sandbox, 2 vCPU and 4 GB | $0.022 per hour | Billed per second; exactly twice the smaller sandbox |
| Desktops and phones | From $0.12 per hour | Linux and Windows desktops, Android phones |
| Growth plan | $100 per month | Becomes $100 of usage; 25 machines awake at once |
| Scale plan | $500 per month | Becomes $500 of usage; 60 machines awake at once |
The per-second billing is what makes the sleep-wake model pay. A 2 vCPU sandbox left running for a full 30-day month costs 720 hours times $0.022, which is $15.84, while the same sandbox used for the typical agent's 5.5 awake hours a month costs about $0.12.
The plan math is just as plain. Growth gives $100 of usage across 25 machines awake at once, which is $4 of included usage per machine, and Scale gives $500 across 60 machines, or about $8.33 per machine. Scale allows 2.4 times as many machines awake at once as Growth, for five times the money, so the bigger plan buys headroom, not a discount.
A desktop or phone at $0.12 an hour costs $2.88 for a full day of use, and a student with a school email can apply for $100 of Growth usage or $500 of Scale usage free each month. Usage-based billing raises the accounting question of what counts as billable work and who can verify it, which is exactly what What IoA Usage Accounting is about.
One warning about the numbers. No dedicated pricing page was found for this article, so every figure above comes from the vendor's blog, its company page, or its students page, all checked on 6 November 2026. Prices for early products move, so confirm the current figures on maritime.sh before you budget a fleet.
The $46 vs $2 Lesson: Polling vs Event-Driven Agents
Maritime's own cost math is the clearest argument for its model, and it comes from a real product the founders run. Their open-source personal agent, Instinct, wakes up on a schedule to check for work, and the vendor priced both ways of running it in a September 2026 blog post.
An agent that wakes up every 30 minutes to check for work costs about $46 per user per month. One that wakes on events costs $2. That is a 23-fold difference, 46 divided by 2, for the same agent doing the same job, and the only change is how it decides to wake up.
| Agent pattern | Monthly cost per user | Why |
|---|---|---|
| Wakes every 30 minutes to poll | About $46 | 48 wake-ups a day, most finding nothing to do |
| Wakes on events only | About $2 | Sleeps until a message, webhook, or cron trigger arrives |
| Always-on sandbox, 2 vCPU | About $15.84 | 720 hours at $0.022, the price of never sleeping |
| Typical agent, 5.5 awake hours | About $0.12 | Sleep-wake billing on the same sandbox rate |
Polling every 30 minutes means 48 wake-ups a day, and most of them find nothing to do. Event-driven waking means the agent sleeps until a message, a webhook, or a cron trigger arrives, and the vendor's post lists four fixes for the polling bill, starting with webhook triggers instead of a timer.
Counting what a run costs is a separate problem from paying for it. What Agent Run Metrics is covers the measurement side, the events and dimensions you log so a bill can be checked, while Maritime's pricing covers the paying side.
Maritime vs Render Sandboxes: Persistent Home vs Throwaway Room
Maritime vs Render sandboxes is the comparison that decides which product fits, because the two solve different halves of the same problem. Render sandboxes are ephemeral cloud sandboxes: a room you rent for one task and leave when it ends. Maritime agent boxes are persistent homes: the agent lives there, keeps its things there, and sleeps there between jobs.
| Question | Maritime agent box | Render sandbox |
|---|---|---|
| Does it persist? | Yes, disk, memory, files, and logins survive | No, built for one task, then gone |
| What wakes it? | A message, webhook, or cron trigger | A new task, from scratch |
| What is it for? | Assistants and agents with a memory | Code execution and single-shot work |
| What does idle cost? | Storage only while parked, $1 a month a box | Nothing, because nothing is kept |
| Who owns the state? | The agent, on its own disk | The caller, who must store results elsewhere |
Use both and the split is natural. Maritime's own sandboxes play the Render role inside Maritime: throwaway code runs in a sandbox, and the repo lives in a workspace that is still there next week. The persistent box holds the memory, and the sandbox does the dangerous work.
The full Render picture is in What Are Render Sandboxes, which covers agent sandboxes on Render in the same beginner's-guide style. Read it before you choose, because the wrong shape costs either money on idle machines or time rebuilding state.
Maritime vs a Plain VPS
A plain VPS is the default home for a self-hosted agent, and the comparison is honest, not a straw man. A VPS gives you a real machine, root access, any software, and a bill that never sleeps, and for one agent that you watch closely it is often the cheapest option.
| Question | Maritime agent box | Plain VPS |
|---|---|---|
| Idle cost | Storage only while parked | Full hourly or monthly price, awake or not |
| State between tasks | Built in, on the agent's own disk | Built in, on the server's disk |
| Wake on message | Built in, about a second | You build it, or leave the server on |
| Per-agent isolation | One microVM per agent | You build it with users, containers, or VMs |
| Scaling to 100 agents | 100 boxes, one API call each | 100 servers, or one server with 100 tenants |
| Ops work | None, the platform runs the hosts | Yours: patching, disk, restarts, monitoring |
The VPS wins on control and on price for a single busy agent. Maritime wins when the fleet grows, because 100 agents on one VPS means 100 tenants sharing one disk, one crash domain, and one set of credentials to separate by hand.
The crossover is easy to estimate. One agent box costs $12 a year, so 100 agents cost $1,200 a year before compute, and a single VPS that hosts the same 100 agents costs whatever its monthly price times 12. If the VPS is cheaper and you are willing to do the ops work, keep the VPS, and if the ops work costs more than the difference, the boxes win.
The Home and the Address: Maritime, NANDA Index, and the MIT Conference
Maritime gives an agent a home. NANDA Index gives it an address. That is the vendor's own framing, from a 20 October 2026 blog post by co-founder Maria Gorskikh, and it is the strongest idea in the whole product story.
About 500 people at an MIT conference each got a personal AI agent of their own, according to that post, and every agent ran on its own machine on Maritime, had its own phone number, and could text its owner. The best moments came when the agents started talking to each other: one person asked their agent to find dinner time with someone across the room, and two agents worked it out on their own.
That worked because each agent had been told exactly where the other one lived. In a room of 500 that is fine, and for the billions of agents the founders expect, it does not work at all. A home without an address means nobody can find you, and an address without a home points at nothing.
The NANDA Index paper, by Ramesh Raskar and colleagues at the MIT Media Lab, proposes the address side. It is a preprint for IEEE CIC 2026, and it adds no new protocol: it gives an agent a stable identifier and points that identifier at wherever the agent's card and discovery information live, using the existing AI Catalog and ARD formats. Because the identifier is separate from the runtime, an agent can sleep, wake, move hosts, or change providers and still be reached at the same address.
That separation is the whole point for a platform where agents sleep most of the time. The post was published 17 days before this article, on 20 October 2026, and the conference it describes is unnamed in the post, with no independent confirmation found. Treat the 500-agent story as the vendor's own account, not as an independently reported event.
NANDA registration on Maritime is exploratory, not a shipped feature. The company says it is exploring how agents deployed on Maritime can be registered with NANDA Index as part of creating them, so a person or small business gets an agent and an address in one step, but that step does not exist yet.
Who Maritime Fits — and Who Should Wait
Maritime fits teams whose agents need memory, and it fits less well when the requirements are audit papers rather than working agents. The product is young, launched in 2026, and the honest fit check is about what you can verify today.
| You should try Maritime if | You should wait if |
|---|---|
| You run personal assistants that need their own files and logins | You need published uptime figures or a service-level agreement |
| You deploy one agent per customer and want isolation by default | You need security certifications, none of which were found in the sources for this article |
| Your coding agent needs a sandbox and a repo that survives | Your data cannot leave a named region, and no region list was published |
| Your evals need hundreds of sandboxes for one job | You need NANDA registration, which is exploratory, not shipped |
| Your agent drives a desktop or mobile app with no API | You want a settled signup flow, and the beta and waitlist signals still conflict |
Students get the easiest start. A school email unlocks $100 of Growth usage or $500 of Scale usage free each month, which is enough to run a real assistant and a swarm or two without touching a card. Three free agent boxes are the same offer for everyone else.
The fleet case is where the model earns its keep. One agent per customer means each customer gets isolation without a DevOps project, and the per-agent price makes the cost of the hundredth customer as predictable as the cost of the first. The risk case is the reverse: if your agents need compliance paperwork, the sources for this article contain no certifications, no region list, and no uptime history to show an auditor.
What to Check Before You Deploy
Check the signup flow live before you plan around it. The March 2026 Product Hunt post described a beta with a manual waitlist, while the November 2026 posts present Maritime as live with three agents free, so check the signup page yourself before assuming either.
Check the price on the day you deploy. No dedicated pricing page was found for this article, so the figures here come from the vendor's blog, company page, and students page as of 6 November 2026. Early products reprice, and a fleet budget built on a blog post is a guess.
Maritime is young. The first Product Hunt launch was 21 March 2026, which is 230 days before this article, and the company joined the Y Combinator Fall 2026 batch the same year.
Check what the free tier actually includes. Three agents free is the headline, and the limits on compute size, sleep behaviour, or expiry are not documented in the sources fetched for this article. Run one agent for a week and read the bill before you move a fleet.
Check the name before you search or invest. Maritime Fusion, a fusion-energy company, is a different company from Maritime (maritime.sh), and the two share nothing but a word. The same care applies to the shipping industry, which owns the bare word Maritime in search results.
CodexFuse does not offer a Maritime tool. This article is a beginner's guide to the product, not a listing for one, and nothing here promises a CodexFuse version of it.
Sources
Maritime homepage, "Every computer your agent needs" — https://maritime.sh/
Maritime blog, Maria Gorskikh, "Every agent needs a home and an address", 20 October 2026 — https://maritime.sh/blog/nandaindex-every-agent-needs-a-home-and-an-address
Maritime blog index, including "I priced Instinct's proactive agents", 29 September 2026, and the snapshot-restore paper, 3 September 2026 — https://maritime.sh/blog
Maritime for Students, Growth and Scale plan pricing — https://maritime.sh/students
LinkedIn, Maritime (YC F26) company page — https://www.linkedin.com/company/maritimeai
Product Hunt discussion, "What kind of computer does your AI agent need?", published 2 November 2026, updated 3 November 2026 — https://www.producthunt.com/p/maritime/what-kind-of-computer-does-your-ai-agent-need
Product Hunt launch, Maritime, 21 March 2026 — https://www.producthunt.com/p/maritime/maritime
NANDA Index paper, Raskar et al., MIT Media Lab, preprint for IEEE CIC 2026 — https://nandaindex.org/paper.pdf
Extruct AI, Y Combinator Fall 2026 companies list — https://www.extruct.ai/data-room/ycombinator-companies-f26/
PyPI, maritime Python SDK, version 0.7.1 — https://pypi.org/project/maritime/0.7.1/
Frequently Asked Questions
What is Maritime AI agent hosting?
Maritime AI agent hosting means giving an AI agent its own persistent cloud computer on Maritime (maritime.sh), a Firecracker microVM with its own disk, memory, files, and logins. The machine sleeps when idle and wakes in about a second when a message arrives, and an agent box costs $1 per agent per month, with three agents free.
How much does Maritime cost?
Agent boxes cost $1 per agent per month, with three agents free, and sandboxes cost $0.011 an hour for 1 vCPU and 2 GB or $0.022 an hour for 2 vCPU and 4 GB, billed per second. Desktops and phones start at $0.12 an hour, and the Growth plan is $100 a month while the Scale plan is $500 a month, with each plan fee becoming the same amount of usage.
Is Maritime a shipping company?
No. Maritime (maritime.sh) is a cloud company for AI agents, and it has nothing to do with ships, ports, or maritime shipping. The bare word Maritime mostly returns shipping news in search results, so always search with the qualified name or the maritime.sh domain.
How is Maritime different from Render sandboxes?
Maritime agent boxes are persistent homes where an agent keeps its disk, files, and logins between tasks, while Render sandboxes are ephemeral rooms built for one task and gone when it ends. Use a Maritime box when the agent needs memory, and a sandbox when the work is throwaway code execution or a single-shot job.
Can I run Claude Code or OpenClaw on Maritime?
Yes. Maritime offers templates for Claude, OpenClaw, Hermes, and OpenAI agents, and you can also bring your own framework by binding the port Maritime gives you and answering a health check and chat messages. Your model key is stored encrypted and injected only into your agent's own VM, and you can drive the computer through MCP from an agent you already run elsewhere.
