Most cloud platforms were built assuming a human is the one clicking around a dashboard, choosing a region, and wiring up a database. That assumption is starting to break: when an AI coding agent is the one doing the work, it needs to provision infrastructure through an API or a command line, not a UI it can't see. A small wave of platforms has shown up specifically for this — built so an AI agent can spin up a database, a sandbox, or a whole backend on its own. Here are five real ones, each solving a different piece of that problem, with honest pricing and where they fall short.
Box: the cheap, no-frills VM for agents to run code in
Box gives an AI coding agent a real, persistent Ubuntu virtual machine it can use to run code, install software, and reset or copy instantly — not a limited container. It positions itself explicitly as the cheapest option against E2B, GitHub Codespaces and Modal, at a $20/month minimum that converts into roughly 555 hours of 4 vCPU/8GB compute, billed by the second past that. It supports disk-level snapshot forking, so an agent can try something risky on a copy without touching the original. The real limits: it's currently only available in Europe (Germany, Finland, France), it's a closed product built on a broader platform (ASCII) whose long-term backing is worth watching, and it's competing against better-funded players who could cut prices.
Specific: a full backend stack an agent can provision itself
Specific is built so an AI assistant — not a person — can provision a database, storage and APIs directly, with a native MCP server as the primary interface. It bundles a genuinely complete stack: Postgres with point-in-time recovery, object storage, Redis, a vector store, and workflow support, plus edge deployment with CDN and TLS included. Pricing starts free (5 projects, 512MB database), then $25/month on Pro and $299/month on Scale, both with usage on top. The trade-offs are the usual ones for a young platform: usage-based pricing is harder to predict than a flat fee, the free tier is fairly limited, and its community and ecosystem are still much smaller than an incumbent like Supabase.
CreateOS: the governance layer for running agents in actual production
CreateOS is aimed at a different, later problem: getting an AI agent from a working demo into production with real governance — kill-switches, autonomy thresholds, audit trails — that an enterprise security team will actually sign off on. It's SOC 2 Type II and ISO 27001 certified, supports air-gapped deployment, and lists solid named customers including Stripe, AWS and Supabase. Pricing is a 2.5% platform fee on top of the lowest model rate plus per-second sandbox billing, with no egress fees or monthly minimum — but real-world cost is hard to estimate without published invoice examples, the onboarding methodology assumes a 12-week commitment, and all of this is likely overkill if you're a small team without a security review process to satisfy in the first place.
InsForge: an open-source backend an agent can clone and branch
InsForge gives a coding agent a complete backend — database, auth, storage, functions — provisioned through a CLI built for agent use, with a genuinely useful trick: git-style branching that lets an agent clone an isolated full copy of the backend to experiment in without risking the original. It's open source (Apache 2.0) with around 12,400 GitHub stars, self-hostable for free, and ships with Stripe and PostHog integrations ready to go. Worth knowing before you commit: there's no published pricing for any future hosted offering, its "agent-native" positioning is partly marketing on top of what is, underneath, a fairly conventional backend, and the company hasn't disclosed its funding or founding date.
Manufact: hosting built specifically for MCP servers
Manufact targets a narrower, newer need: hosting for MCP servers — the plumbing that lets an AI assistant like Claude or ChatGPT actually call your tools and data. Instead of wiring that hosting together yourself on a generic VPS, Manufact handles deploy, monitoring and scaling for the MCP lifecycle specifically, and it's built on the open-source mcp-use SDK, so self-hosting stays an option and you're not fully locked in. Pricing is clear and tiered: free (5 credits, 2 projects, 7-day retention), Hobby at $25/month, Startup at $250/month, and custom Enterprise. The jump from Hobby to Startup is steep, MCP hosting is a brand-new category so the product and pricing will likely keep shifting, and self-hosting the open SDK remains a valid alternative for anyone who'd rather avoid a dedicated platform altogether.
| Platform | What it's for | Price | Best fit |
|---|---|---|---|
| Box | Cheap persistent VM to run code in | $20/mo min (~555h compute) | Cost-sensitive teams in Europe |
| Specific | Full DB/storage/Redis/vector stack, agent-provisioned | Free / $25 / $299 + usage | Agent-built apps needing a real backend |
| CreateOS | Governance & kill-switches for production agents | 2.5% platform fee + sandbox usage | Enterprises with a security review to pass |
| InsForge | Open-source backend agents can clone/branch | Free (self-hosted), hosted pricing TBD | Teams wanting no lock-in |
| Manufact | Hosting specifically for MCP servers | Free / $25 / $250 + custom | Anyone shipping an MCP integration |
None of these compete head-on — they sit at different layers of the same emerging stack. If you just need somewhere cheap for an agent to run code, start with Box. If you need a real backend an agent can provision and branch on its own, Specific and InsForge cover the hosted and self-hosted ends of that. If you're past the demo stage and need a security team to approve putting an agent into production, CreateOS is built for exactly that conversation. And if your specific problem is hosting an MCP server, Manufact is the only tool here built for that job alone.