RetroDeckRETRODECK

Maritime Launches a Managed Cloud for Production AI Agents: What It Offers and What HN Asked

Maritime, a Y Combinator Fall 2026 startup that launched on Hacker News on November 18, 2026, runs each AI agent in its own persistent Firecracker microVM and manages sleep, storage, triggers, spend caps and updates around it. Agent Boxes start at $1 a month. Commenters questioned crowding, pricing, patching and restore behaviour.

What Maritime launched

Maritime, a Y Combinator Fall 2026 company founded by Ary and Maria, posted its Launch HN on November 18, 2026. The pitch is that sandboxes solve only a small part of the problem of running agents for customers. After picking a sandbox provider, the founders wrote, teams still build sleep and wake, persistent storage, triggers, scheduling, per-user routing and spend caps, health checks, failure recovery and rolling updates across thousands of VMs.

Maritime's answer is one isolated, persistent microVM per agent, with those lifecycle services built around it. Its site lists Agent Boxes (persistent machines with a disk at /data that sleep when idle and wake in about a second), single-task Sandboxes, Swarms for evals and batch work, Linux and Windows desktops for computer-use agents, and Android 15 phones. Agents can be OpenClaw, Hermes, the Claude Agent SDK, Codex or any Docker image that listens on $PORT, exposes /health and /chat, and keeps state under /data. Control is through a REST API, TypeScript and Python SDKs, a CLI and MCP.

Pricing, as published

The headline is $1 a month for a 2GB Agent Box, billed awake or asleep, with unlimited awake hours and 5GB of disk; a team member on the thread put the box at 1 vCPU. Larger boxes are $2 for 4GB and $4 for 8GB. Plans are Starter at $20, Growth at $100, Scale at $500 and Dedicated at $5,000 a month, each converted into usage credit, and they differ mainly in how many machines can be awake at once: 5, 25, 60 and 500. Sandboxes bill per second at $0.0055 per vCPU-hour plus $0.00275 per GB-hour. Desktops and phones bill while awake: Linux $0.12, Windows $0.18 and Android $0.15 an hour. Model inference is extra.

The founders' own example explains the design. A sandbox at $0.04 an hour, left running, is about $29 a month, more than a $20 personal assistant earns. Scale-to-zero, copy-on-write base images, compressed memory snapshots and tiered storage are how Maritime claims to make an always-available agent cost $1.

How it compares with other agent runtimes

Agent runtimes split into two shapes, and Maritime sits firmly in one of them.

Framework runtimes

Google's Agent Runtime and Amazon's Bedrock AgentCore Runtime host agent code. You deploy a LangGraph, ADK or custom agent, and the platform manages scaling, sessions, memory and tracing around it. The unit is an agent invocation inside a cloud you already use, and the runtime is tightly bound to that provider's identity, models and observability.

Machine-per-agent runtimes

Maritime's unit is a computer: a persistent VM with its own kernel, disk and network, which is what computer-use, browser and coding agents need. The closest comparisons commenters raised were DigitalOcean's managed agents product, Agent37, Daytona, Playcode's cloud and Kubernetes options such as kagent. A founder replied that kagent is a framework you run on your own cluster while Maritime runs the infrastructure, and that CNCF's agent-substrate shares workers between many agents where Maritime gives each its own VM.

What is genuinely different

Maritime's security page says sandboxes and Linux desktops are Firecracker microVMs hosted at Hetzner in Falkenstein, Germany, with Windows desktops on AWS us-east-1. That is the commodity layer the launch post criticised other providers for selling, so the differentiation is the managed lifecycle on top: snapshots, triggers, spend caps, updates and patching. That is a reasonable thing to buy, but it should be judged on operations, not on the VM.

The substantive objections on Hacker News

The thread reached 58 points within a day, and the pushback was concrete. The first was crowding: several commenters listed near-identical offerings, and one linked a directory of harnesses, sandboxes and platforms doing the same job. Another asked how the company recoups its funding while charging $1 a box; the founder answered that sandboxes are a small part of the offering and revenue scales with agent count.

Security drew the most experienced comment. An engineer at a neocloud wrote that bringing up Firecracker on bare metal is easy, and the real work is continuous patching of host and guest kernels against new CVEs, which argues against picking a small provider on price. The founder said patching is part of what managed means. Compliance is partway there: the security page lists a HIPAA attestation issued October 19, 2026 and describes SOC 2 as underway.

Two technical questions stood out. On spend caps, Maritime returns HTTP 429 from its LLM gateway when a cap is hit, and loop detection is left to your harness. On sleep and wake, in-progress tool calls survive but open connections do not. A follow-up asked what happens on an unplanned restore from snapshot, when an agent may repeat a push, message or charge it already made, and whether the VM is told it was restored. That question had no answer in the thread when checked on November 19. Separately, one commenter said that using consumer Claude or ChatGPT subscriptions inside third-party tools breaks those providers' terms, so check the terms and plan on API keys.

What to check before moving agents onto it

Treat restore semantics as the first question. If your agents take actions with side effects, make them idempotent with your own ledger of completed actions, whichever host you use. Next, check data residency against your customers: compute sits in Germany and the control plane in the US. Then price your real wake pattern, since Windows desktops cost far more per hour than Agent Boxes.

Finally, a VM per customer solves compute, not knowledge. Thousands of isolated boxes each need the same product docs, policies and account context, and copying that into every /data volume is the same sync problem the founders describe for storage. RDK is built as that shared layer: agents in any runtime query it through MCP, the CLI or the Partner API and get a few hundred tokens of context, while private chunk content stays on your nodes, encrypted with the vault key, and team redaction runs at egress.

Frequently asked questions

What is Maritime?
Maritime is a managed cloud for production AI agents from a Y Combinator Fall 2026 company, launched on Hacker News on November 18, 2026. It gives each agent its own persistent Firecracker microVM and manages sleep and wake, storage, triggers, spend caps, updates and recovery. It also offers sandboxes, Linux and Windows desktops and Android phones.
How much does Maritime cost?
Its site lists $1 a month for a 2GB Agent Box, billed awake or asleep, with plans at $20, $100, $500 and $5,000 a month that convert into usage credit. Plans differ in how many machines can be awake at once. Sandboxes, desktops and phones are billed by usage, and model inference costs extra.
How is Maritime different from Bedrock AgentCore or Google's Agent Runtime?
AgentCore Runtime and Google's Agent Runtime host agent code inside their clouds and manage sessions, memory and tracing. Maritime gives each agent a whole machine, with its own kernel, disk and network, including Windows and Android environments. That suits computer-use and coding agents, while the cloud runtimes suit agents built around one provider's stack.
Where does Maritime run agents?
According to its security page, sandboxes, Linux desktops and Android phones run in Germany at Hetzner in Falkenstein, Windows desktops run in AWS us-east-1, and the control plane and API are in the United States. Enterprise customers can add dedicated hosts. Check this against any data residency commitments you have made to customers.