Compute & sandboxes · Decision brief
Alfe vs Fly.io Machines
The finished agent OS on top of the micro-VM.
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.
Alfe
Managed agent OS
Best when you need
A persistent agent with compute, models, memory, identity and channels managed together.
Fly.io Machines
Compute & sandboxes
Best when you need
Low-level, globally placed micro-VMs.
The short answer
Which one fits the job?
These products often sit at different layers of the stack. The useful question is not which has more ticks—it is whether you want to assemble the system or operate a finished agent.
Choose Alfe when
A place to run code vs a finished agent
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.
Pooled model billing, not a second bill
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.
Managed memory vs Fly Volumes
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.
Choose Fly.io Machines when
- Global edge reach & subsecond VMs: Subsecond start/stop micro-VMs across 18 regions on six continents — Fly's core strength
- Low-level control & flexibility: Raw, low-level micro-VMs you operate directly — Fly's core strength for custom runtimes
Capability matrix
Alfe and Fly.io Machines, side by side.
A practical comparison of product shape, operations and the capabilities a team receives without additional assembly.
| Capability | Alfe | Fly.io Machines |
|---|---|---|
| What it is | AlfeA managed agent OS — runtime, models, memory, channels and identity in one platform | Fly.io MachinesHardware-virtualized micro-VMs driven by a REST API / flyctl, commonly used to host agents & MCP servers |
| Out-of-the-box agent runtime | AlfeOpenClaw + Hermes on a dedicated per-agent server, managed lifecycle + crash recovery | Fly.io MachinesNone — Fly boots a micro-VM; you supply the agent runtime and orchestration |
| Model access & billing | AlfePooled proxy across 9 providers metered into one prepaid USD credit pool | Fly.io MachinesNot applicable — pure compute; you bring your model and pay it separately |
| Managed agent memory | AlfeSemantic vector store + a knowledge graph, managed and persistent | Fly.io MachinesDurable state via Fly Volumes + snapshots; Machine RAM is ephemeral — infra state, not agent memory |
| Global edge reach & subsecond VMs | AlfeDedicated per-agent server (Hetzner/ECS) — a stable home, not an 18-region edge fabric | Fly.io MachinesSubsecond start/stop micro-VMs across 18 regions on six continents — Fly's core strength |
| MCP self-bootstrap | AlfeNative MCP + agents self-onboard over mcp.alfe.ai (proof-of-work → claim own compute + identity) | Fly.io MachinesNative `fly mcp` hosts a remote MCP server, but agent self-provisioning is yours to build |
| Channels | AlfeSlack, Discord, Teams, Google Chat, web, mobile — plus voice, SMS & WhatsApp on a phone number | Fly.io MachinesNone — Fly is compute; channels are yours to build |
| Voice & phone | AlfeStreaming voice, SMS, and WhatsApp on a real number | Fly.io MachinesNot offered |
| Teams, orgs, fleets & identity | AlfeFull org hierarchy + OAuth-provisioned per-agent bots and credentials | Fly.io MachinesFly orgs for the compute account; agent identity and roles are yours to build |
| Low-level control & flexibility | AlfeAn opinionated managed agent — less to configure, less to control at the VM layer | Fly.io MachinesRaw, low-level micro-VMs you operate directly — Fly's core strength for custom runtimes |
Where Alfe differs
The operating layer is the product.
A place to run code vs a finished agent
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.
Pooled model billing, not a second bill
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.
Managed memory vs Fly Volumes
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.
Where Fly is stronger
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.
Channels, voice and teams out of the box
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.
How this comparison is made
Transparent by design.
This is a first-party Alfe comparison, not an independent review. We assess product positioning, hosting, model billing, memory, MCP, team controls and channels against publicly available product information. Products change; use the official source below for the latest detail.
Read Fly Machines documentation Compared with Alfe · Fly.io MachinesQuestions teams ask
Alfe vs Fly.io Machines FAQ.
Is Alfe a Fly.io Machines alternative?
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.
Can Fly host an agent or an MCP server?
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.
Does Fly give me managed agent memory?
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 is Fly the better choice?
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.
Ready to run
Start with the agent, not an empty VM.
Get a managed runtime, pooled model access on one credit pool, managed memory, teams, 40+ integrations and voice — no micro-VM to configure first.