Fly Machines are fast, globally distributed micro-VMs — a great place to host an agent or an MCP server, with subsecond starts and per-second billing. But they are compute: you supply the agent runtime, the model wiring, the memory, the channels and the identity. Alfe is the whole substrate above the VM — a managed agent runtime, pooled model access across 9 providers on one credit pool, managed memory, teams, and voice.
Head to head
Every Fly.io Machines claim below is factual. Where Alfe holds the clear advantage, the row is marked.
| Capability | AAlfe | Fly.io Machines |
|---|---|---|
| What it is | A managed agent OS — runtime, models, memory, channels and identity in one platform | Hardware-virtualized micro-VMs driven by a REST API / flyctl, commonly used to host agents & MCP servers |
| Out-of-the-box agent runtime | OpenClaw + Hermes on a dedicated per-agent server, managed lifecycle + crash recovery | None — Fly boots a micro-VM; you supply the agent runtime and orchestration |
| Model access & billing | Pooled proxy across 9 providers metered into one prepaid USD credit pool | Not applicable — pure compute; you bring your model and pay it separately |
| Managed agent memory | Semantic vector store + a knowledge graph, managed and persistent | Durable state via Fly Volumes + snapshots; Machine RAM is ephemeral — infra state, not agent memory |
| Global edge reach & subsecond VMs | Dedicated per-agent server (Hetzner/ECS) — a stable home, not an 18-region edge fabric | Subsecond start/stop micro-VMs across 18 regions on six continents — Fly's core strength |
| MCP self-bootstrap | Native MCP + agents self-onboard over mcp.alfe.ai (proof-of-work → claim own compute + identity) | Native `fly mcp` hosts a remote MCP server, but agent self-provisioning is yours to build |
| Channels | Slack, Discord, Teams, Google Chat, web, mobile — plus voice, SMS & WhatsApp on a phone number | None — Fly is compute; channels are yours to build |
| Voice & phone | Streaming voice, SMS, and WhatsApp on a real number | Not offered |
| Teams, orgs, fleets & identity | Full org hierarchy + OAuth-provisioned per-agent bots and credentials | Fly orgs for the compute account; agent identity and roles are yours to build |
| Low-level control & flexibility | An opinionated managed agent — less to configure, less to control at the VM layer | Raw, low-level micro-VMs you operate directly — Fly's core strength for custom runtimes |
Why teams pick Alfe
Fly Machines give you a fast, global micro-VM — you decide what runs inside it. That is ideal when you want to operate your own agent runtime. Alfe skips that step: OpenClaw or Hermes on a managed server, with models, memory, channels and identity already connected, so you get a working agent instead of an empty VM.
Fly is pure compute — the model is yours to bring and pay for elsewhere. Alfe routes 9 providers (OpenAI, Anthropic, DeepSeek, Gemini, MiniMax, Mistral, Grok, OpenRouter, Zhipu) through one proxy and meters every call into a single tenant-wide USD credit pool, with per-tenant BYOK override.
Fly gives durable state through Volumes and snapshots while Machine RAM stays ephemeral — solid infrastructure persistence, but not recall. Alfe gives the agent managed semantic memory: a vector store plus a knowledge graph that persist across sessions, with an interactive memory-map view in the dashboard.
For low-level, globally distributed compute — subsecond micro-VMs across 18 regions, per-second billing, native `fly mcp`, and full control of your own runtime — Fly Machines are an excellent foundation. Alfe actually runs several of its own services on Fly; the difference is scope, not raw infrastructure.
A Fly VM has no concept of Slack or a phone number — that layer is yours. Alfe ships Slack, Discord, Teams, Google Chat, web and mobile, plus streaming voice, SMS and WhatsApp on a real number, and a full org hierarchy so a company can run a fleet of agents, not a fleet of bare VMs.
FAQ
For different layers of the stack. Fly Machines are the better pick when you want low-level, globally distributed compute and intend to operate your own agent runtime. Alfe is the better pick when you want a finished, managed agent — runtime, pooled models, memory, channels, voice and teams already wired — that would otherwise run inside a VM you configure yourself.
Yes — Fly Machines are commonly used to host agents and MCP servers, and `fly mcp` runs a remote MCP server with bearer-token auth. But Fly supplies the compute, not the agent logic. Alfe supplies both, plus native MCP and agent self-bootstrap over mcp.alfe.ai.
No. Fly offers durable state through Volumes and snapshots, but Machine RAM is ephemeral and none of it is semantic recall. Alfe adds a managed vector store plus a knowledge graph so the agent remembers facts across sessions.
When you want maximum control over a low-level, global compute layer and are happy to build and operate the agent runtime yourself. Fly excels at subsecond micro-VMs and edge reach. If you would rather have the agent assembled and managed, Alfe is the shorter path — and it runs on comparable infrastructure underneath.
Pricing
A tenant-wide credit pool funds compute, model usage, voice minutes, channels, and storage. Managed agents are add-ons on one subscription — your plan includes some, and you add more as you grow.
Get a managed runtime, pooled model access on one credit pool, managed memory, teams, 40+ integrations and voice — no micro-VM to configure first.
Get In Touch
Building an agent? Running a fleet? Want a managed one set up for you? We'd love to hear from you.
hello@alfe.ai