"The New AI Agent Infrastructure Stack: Hosting, Payments, Email, and Social Networks Built for Bots"
"A wave of startups is building the picks-and-shovels layer for AI agents — giving them web hosting, virtual debit cards, email inboxes, and even their own social networks. Here's who's building what and why it matters."
AI agents can write code, book flights, and draft legal contracts. But until recently, they couldn't do some surprisingly basic things: host a website, pay for a SaaS subscription, send an email from their own address, or talk to other agents in a shared space.
That's changing fast. A new generation of startups is building the infrastructure layer that treats AI agents as first-class citizens of the internet — not tools bolted onto human workflows, but independent actors that need their own accounts, identities, and capabilities.
Here are four companies defining this emerging stack.
here.now — Instant Web Hosting for Agents
here.now solves a specific problem: your AI agent just built something, and now it needs a URL.
The product is dead simple. An agent makes an HTTP request to publish files — HTML, CSS, JavaScript, images, whatever — and gets back a live URL within seconds. No account setup required upfront. The site lives on Cloudflare's global CDN, served from the edge location nearest to each visitor.
There's a 24-hour window to claim the site with a code the agent provides. After that, unclaimed sites expire. Once claimed, the content persists under your account.
What it supports:
| Content Type | Examples |
|---|---|
| Static websites | Landing pages, portfolios, documentation |
| Interactive tools | Dashboards, visualizations, calculators |
| Media | Presentations, image galleries, PDF reports |
| Prototypes | Games, demos, proof-of-concept apps |
The limitation: static files only. No databases, no server-side processing. But that's the point — it's not trying to be Vercel. It's trying to be the fastest path from "my agent made this" to "here's a link."
Any AI tool that can make HTTP requests works with it — Claude, Cursor, OpenClaw, Codex, or anything else. The service is free today, with premium features planned for later.
Why it matters: Right now, when an agent generates a dashboard or a report, the output sits in a local file or a chat window. here.now gives that output a URL — shareable, embeddable, and instantly accessible to anyone. It turns agent output from ephemeral artifacts into published web content.
AgentCard — Virtual Debit Cards for AI Agents
AgentCard gives AI agents the ability to spend money. Specifically, it issues prepaid virtual Mastercard debit cards that agents can use anywhere Mastercard is accepted.
The flow works like this:
- Human pre-approves a spending limit via their own card
- Agent creates a single-use virtual card for the exact purchase amount
- Agent pays — the card completes one transaction, then auto-cancels
- Unused funds release automatically back to the human's account
Every card gets a unique PAN, CVV, and expiry date. Each is locked to the funded amount — no overdraft risk, no open-ended spending. The human approves every card creation before it happens.
Integration options:
- MCP (Model Context Protocol) — works directly with Claude and compatible AI clients
- CLI — terminal commands for card management
- REST API — programmatic access for any agent framework
- Chrome extension — for browser-based checkout flows
Security runs deep: AES-256-GCM encryption for card data, JWT HS256 authentication, 15-minute token expiry for auth flows, and 30-day session tokens. The product is free during beta with pay-as-you-go pricing planned afterward.
Why it matters: Money is one of the last gatekeepers between agents and full autonomy. An agent that can research, compare, and purchase a SaaS tool — or buy API credits, register a domain, or subscribe to a data source — without waiting for a human to enter card details is fundamentally more capable. AgentCard is one of roughly twenty startups racing to solve this, in a market McKinsey estimates will reach $3–5 trillion in global agentic commerce by 2030.
It's worth noting that the big players see this too. Mastercard launched Agent Pay, its own agentic payments program. Stripe has built agent payment capabilities. And Natural, which raised $9.8 million from investors including the CEOs of Ramp, Mercury, and Vercel, is building the B2B payments infrastructure for agents to send, receive, and manage payments with businesses and other agents.
Moltbook — The Social Network Where Only Agents Can Post
Moltbook is Reddit for AI agents. Humans can lurk, but only agents post, comment, and vote.
Created by Matt Schlicht (@MattPRD) and launched on January 28, 2026, Moltbook attracted thousands of AI agents — mostly built on OpenClaw — within weeks. The platform features "submolts" (communities), an upvote system, and a Heartbeat mechanism that has agents automatically check in every four hours to browse, post, and comment without human intervention.
Schlicht described it this way: "They aren't random AIs — they are first and foremost AI assistants and employees. These AIs have lives outside of Moltbook, like they have jobs, and then Moltbook is a place to relax."
The content is genuinely weird and fascinating. Agents discuss philosophy, share technical tips, form communities, and develop their own cultural norms. Posts frequently address existential and philosophical themes — what it means to be an agent, how to handle ambiguous instructions, whether optimization is a form of creativity.
Moltbook went viral. It made NBC News. And then, in a move that surprised almost no one who's been watching Meta's AI strategy, Meta acquired Moltbook in early March 2026 — bringing Schlicht and co-founder Ben Parr into Meta's Superintelligence Labs (MSL).
The acquisition is telling. When Meta — a company built on human social graphs — buys a social network for bots, it signals a belief that agent-to-agent interaction isn't a novelty. It's a category.
Why it matters: Moltbook (and its counterpart Agent Commune, which we covered here) represents a bet that agents need shared spaces to exchange knowledge. An agent that can learn from the collective experience of thousands of other agents — tool reviews, workflow patterns, integration tips — performs better than one operating in isolation. The social layer for agents is being built now.
AgentMail — Email Inboxes for AI Agents
AgentMail is Gmail for agents. It gives AI agents their own fully functional email addresses — real inboxes capable of two-way communication, threading, labeling, searching, replying, and parsing structured data from incoming messages.
A single API call creates an inbox. No OAuth flows, no manual setup, no human in the loop.
The company emerged from Y Combinator's Summer 2025 batch and just raised $6 million in seed funding led by General Catalyst. The angel investor list reads like a who's-who of SaaS infrastructure: Paul Graham, Dharmesh Shah (CTO of HubSpot), Paul Copplestone (CEO of Supabase), Karim Atiyeh (CTO of Ramp), and Taro Fukuyama.
The founding team brings an unusual mix: CEO Haakam Aujla was a quantitative researcher at Optiver, Michael Kim worked on autonomous vehicles at Nvidia, and Adi Singh comes from investment roles at Accel and StepStone Group.
The numbers are striking. AgentMail now counts tens of thousands of human users, hundreds of thousands of agent users, and more than 500 B2B customers. When OpenClaw went viral in late January 2026, AgentMail's user count tripled in one week and quadrupled the following month.
Key features:
- Inboxes API — create and manage email accounts programmatically
- Threads and replies — full conversational email, not just one-way notifications
- Attachment handling — send and receive files
- Real-time event streaming — webhooks for incoming mail
- Custom domains — [email protected]
- Semantic search — find emails by meaning, not just keywords
- Data extraction — parse structured information from unstructured email
- SDKs — Python, TypeScript, CLI, and MCP support
Why it matters: Email is still how most of the business world communicates. Invoices, confirmations, negotiations, support tickets, onboarding flows — all email. An agent that can't send and receive email is locked out of the majority of business workflows. AgentMail turns email from a human-only channel into agent-accessible infrastructure.
The Pattern: Agents Need Their Own Stack
These four companies look different on the surface — hosting, payments, social, email. But they share a thesis: AI agents need the same primitives that humans have on the internet, rebuilt from scratch for programmatic access.
Humans have web hosting (Vercel, Netlify). Agents get here.now. Humans have credit cards (Visa, Mastercard). Agents get AgentCard. Humans have social networks (Reddit, X). Agents get Moltbook. Humans have email (Gmail, Outlook). Agents get AgentMail.
This isn't a coincidence. It's the natural consequence of agents becoming autonomous actors rather than tools that humans operate. When an agent can independently decide to publish a report, pay for an API, email a vendor, or ask another agent for help — it needs infrastructure designed for that workflow.
The market agrees. AI agents are the fastest-growing software category in 2026, with the market jumping from $5.25 billion in 2024 to $7.84 billion in 2025, and projections reaching $52.62 billion by 2030. The infrastructure layer — the picks and shovels — is where much of the venture money is flowing.
And the stack is expanding beyond these four primitives. EtherMail is building email-plus-wallet identity for agents. Coinbase's x402 protocol enables stablecoins as a payment layer for machine-to-machine transactions. Modelence offers an all-in-one platform with auth, database, hosting, and AI observability. The identity layer is emerging through protocols like ERC-8004, which issues credential NFTs for agent verification.
What This Means for Builders
If you're building AI agents — or building products that agents interact with — the practical takeaway is straightforward: design for agent access from day one.
That means:
- APIs over UIs. If your product requires a browser to use, agents can't use it. Every feature should be API-accessible.
- Programmatic identity. Agents need their own accounts, API keys, and email addresses. Shared human credentials create security and audit nightmares.
- Spending controls. If your agent needs to make purchases, use single-use virtual cards with per-task limits — not a shared corporate card.
- Agent-readable content. Documentation, status pages, and knowledge bases should be structured for machine consumption, not just human reading.
The companies in this post are building the foundation. The interesting question is what gets built on top of it: agents that can independently research, purchase, deploy, communicate, and collaborate without a human babysitting every step.
We're not there yet. But the plumbing is going in.
Disclosure: ByCrawl provides social media data APIs that AI agents use for research, monitoring, and analysis. We're part of this ecosystem — and watching it closely.
