Fly.io Platform

Use this skill for any Fly.io work, including routine deploy/config changes and architecture decisions for apps known to run on Fly.

Fly changes quickly. Ground decisions in the local CLI and current official docs before making version-sensitive claims. When a Fly-authored feature is only documented in Fly forum announcements or official example repos, say so and avoid presenting it as broadly documented platform surface.

First Steps

1. Inspect the project shape before prescribing commands.

   rg --files -g 'fly.toml' -g 'Dockerfile*' -g '.dockerignore' -g '.github/workflows/**' -g 'package.json' -g 'pyproject.toml' -g 'go.mod' -g 'Cargo.toml'

1. Check the current flyctl surface for the task.

   fly version
   fly help
   fly <command> --help

1. Prefer non-mutating validation unless the user explicitly asked to deploy or

change remote resources.

   fly config show --local
   fly config validate --strict
   fly status
   fly checks list

1. Use official Fly docs as the source of truth for unstable details. Do not

use arbitrary community posts as encoded facts. For Fly-Src, use the Fly-authored forum announcements and official example repo linked from references/networking.md, and state that limitation when it matters.

Reference Routing

• Read references/deploy-config.md for fly.toml, process groups, health

checks, deploy strategies, GitHub Actions, secrets, scaling, autostop, and config validation.

• Read references/networking.md for 6PN, .internal, Flycast, public

services, private services, custom private networks, request routing headers, fly-replay, Fly-Src request source auth, and static egress IPs.

• Read references/data-storage.md for Managed Postgres, volumes,

auto-extension, snapshots, backups, Tigris, SQLite with Litestream, Upstash Redis, LiteFS, and durability decisions.

• Read references/production.md for production reviews, security, orgs,

tokens, public IP audits, backups, monitoring, logging, and extensions.

Fly-Native Defaults

• For ordinary apps, prefer Fly Launch-managed apps (fly launch, fly deploy,

fly.toml) over hand-managed Machines.

• Use fly machine or the Machines API only when the task needs per-Machine

control, unmanaged workloads, one-app-per-customer isolation, or runtime code execution patterns.

• For production Postgres, prefer Managed Postgres (fly mpg) over legacy

unmanaged Postgres.

• For object storage, prefer Tigris.
• For simple single-writer SQLite apps, use a Fly Volume plus Litestream to

Tigris when restore-based durability is acceptable.

• For Redis, use Fly's Upstash integration unless the app has a clear reason to

run its own Redis.

• For private proxy-routed services, prefer Flycast over raw .internal when

the service needs Fly Proxy behavior such as autostart, load balancing, TLS, PROXY protocol, or DNS-hostile clients.

• For Fly app-to-app HTTP origin checks, use verified Fly-Src metadata when

the request path goes through Fly Proxy and caller identity is the auth boundary. Do not replace user auth or third-party auth with Fly-Src.

• For fixed outbound allowlisting, prefer app-scoped static egress IPs with

fly ips allocate-egress; do not recommend legacy machine-scoped egress IPs for new work.

Gotchas To Surface Early

• Data locality: Machines, volumes, regions, and data durability are not

interchangeable. Volumes are local NVMe slices, not replicated network disks.

• Network exposure: a service in fly.toml plus public IPs exposes that service

publicly, even if the intended consumer is private or Flycast.

• Fly-Src: never trust Fly-Src without verifying Fly-Src-Signature against

/.fly/fly-src.pub, and check timestamp freshness before authorizing.

• CLI surprises: current fly launch --no-deploy can also offer GitHub Actions

setup in GitHub repos; use --no-github-workflow when that automation is not wanted.

• Process groups: defining [processes] makes the list complete. Removing a

process group from config can destroy Machines in that group on deploy.

• Health checks: service-level checks affect routing, but failing checks do not

restart or stop Machines by themselves.

• Secrets: fly secrets set restarts Machines unless staged; secrets are runtime

env vars, not build args.

• Release commands: they run in temporary Machines without attached volumes.
• Egress: default outbound IPs are unstable. Allocate app-scoped egress IPs per

region only when third parties require allowlisting.

• Litestream: it is async backup/restore for SQLite, not HA, synchronous

replication, multi-writer SQLite, or automatic failover.

• LiteFS: use with caution, keep off-site backups, and do not combine LiteFS

with Fly Proxy autostop/autostart.

• GPUs: do not recommend Fly GPUs for new work. Fly docs say GPUs are deprecated

and unavailable after August 1.

Architecture Triggers

Pause for Fly-specific architecture thinking when the prompt involves state, databases, SQLite, volumes, private services, internal APIs, app-to-app auth, regions, tenants, queues, workers, egress allowlisting, HA, failover, restore objectives, production, or cost boundaries. In those cases, answer with:

1. The Fly-native recommendation.
2. The exact commands or fly.toml shape.
3. The gotchas checked and any remaining production risks.

Downrank beginner tutorials, dashboard-first workflows, deprecated offerings, and experimental surfaces unless the user asks directly.