Build Your Own Edge Network on Commodity Hardware

Fly.io lets you scatter micro-VMs across 30+ regions with a single command. It’s genuinely great. It’s also someone else’s hardware with someone else’s pricing — and those micro-VMs come with 256 MB of RAM.

What if you could get the same geographic distribution on servers with 16x the memory, for a similar monthly bill? Three VPS boxes, one tako.toml, and Cloudflare routing users to the nearest one.

The architecture

Three layers. Cloudflare sits at the edge — handles DNS, terminates TLS, and routes each request to the nearest origin. Tako runs on each VPS — deploys your app, manages processes, handles zero-downtime rolling updates. Your servers do the actual compute.

A user in Tokyo hits Singapore. Berlin goes to Frankfurt. San Francisco goes to LA. Each server runs the same app, deployed from the same config, completely independent of the others.

The config

We covered multi-server config in detail in One Config, Many Servers. The short version:

name = "myapp"

[build]
run = "bun run build"

[envs.production]
route = "myapp.com"
servers = ["la", "fra", "sgp"]

Register each server once with tako servers add, then tako deploy builds your app once locally and uploads it to all three servers in parallel via SFTP. Each server runs its own rolling update independently — if Frankfurt finishes before Singapore, it starts serving the new version immediately.

The routing layer

You have two options, depending on how much you want to spend on intelligence.

Free: round-robin DNS. Add three proxied A records for myapp.com in Cloudflare. Requests distribute across your origins, and Cloudflare retries failed ones automatically. You get failover for free, but no geo-awareness — a user in Tokyo might hit LA.

Smart: geo-steering ($20/mo). Cloudflare Load Balancing with geo-steering assigns each origin to a geographic region. Health checks run every 60 seconds — if Singapore goes down, traffic fails over to the next nearest server automatically.

The free option is a fine starting point. Upgrade to geo-steering when latency starts to matter.

The cost

Fly.io wins on simplicity and minimum price. But look at the resources: three Hetzner boxes give you 4 GB per region — 16x Fly.io’s cheapest tier. For $38/month with geo-steering, you get 12 GB of total RAM across three continents on machines you fully control. Without geo-steering, $18/month still gives you three globally distributed servers with automatic failover.

And those VPS boxes can each host multiple apps that scale to zero when idle — your edge network doesn’t have to serve just one project.

No control plane

Each Tako server is self-sufficient. Its own Pingora proxy, its own process management, its own secrets database, its own release history. There’s no cluster state, no leader election, no orchestrator to babysit. If one server disappears, the others keep serving and Cloudflare stops routing to it.

Adding a region later is a three-step process: spin up a VPS, run tako servers add, add the name to your servers list. Next deploy, it’s live. Removing one is the reverse — take it out of the list, decommission the box.

The bigger picture

This is the floor. The same config that deploys to three servers also handles per-server scaling, per-environment secrets, and a staging environment that costs nothing when nobody’s using it.

And Tako is growing past deploys. The SDK is the starting point for backend primitives — WebSocket channels, queues, workflows — running on the same servers, managed from the same config. Three boxes in three regions, each with the full platform layer, all from one tako.toml.

Your edge network doesn’t need to be someone else’s infrastructure. A few cheap VPS boxes and the right tools get you surprisingly far.

Get started →