Fly Machines

Fly.io Machines API exposes the full lifecycle of individual VMs — hardware-isolated microVMs (Firecracker) that start in milliseconds, run any Docker image, and are reachable anywhere via Fly.io's WireGuard mesh. Unlike the high-level `fly deploy` workflow, the Machines API gives per-VM control: create a machine, start it, exec commands in it, stop it, destroy it. For agents needing ephemeral compute: spin up a Machine to run a task, exec the workload, collect output, destroy the Machine — paying only for actual run time. GPU machines available in select regions. Persistent volumes attach to machines for stateful workloads. Global network (35+ regions) minimizes latency for distributed deployments. Confidence is docs-derived.

REST API at api.machines.dev. Resources: apps, machines, volumes. POST /v1/apps/{app}/machines creates a machine ({config: {image, env, services, guest: {cpus, memory_mb, gpu_kind}}}, region). POST /v1/apps/{app}/machines/{id}/start starts a stopped machine. POST /v1/apps/{app}/machines/{id}/stop stops a running machine. DELETE /v1/apps/{app}/machines/{id} destroys a machine. POST /v1/apps/{app}/machines/{id}/exec runs a command in the machine (body: {command: ["/bin/sh", "-c", "echo hello"]}). GET /v1/apps/{app}/machines lists all machines with state. POST /v1/apps/{app}/volumes creates a volume. Machine states: created → started → stopped → destroyed.

Fly API token auth: Authorization: Bearer {fly_token}. Tokens from fly.io Dashboard → Account → Tokens, or CLI: fly tokens create. App-scoped or account-scoped tokens. HTTPS enforced. No OAuth2. App-level access control: members/viewers managed via fly CLI. Private networking: machines within an org communicate via 6PN (private IPv6) WireGuard mesh — no public exposure needed for internal communication. Machine network isolation: each machine gets its own network namespace.

Standard HTTP status codes with JSON error bodies. 400 for invalid machine config (bad image reference, unknown region). Machine start/stop: async — poll GET /machines/{id} for state transition. Exec: synchronous; returns stdout/stderr/exit_code. GPU availability: limited regions and instance types — check availability before creating GPU machines. Volumes: persist across machine stop/start; destroyed with machine if not detached first. Fly.io uptime at status.fly.io. MicroVM startup: typically 200ms–2s (cold start from image). Machine crash detection and automatic restart configurable via restart policy in machine config.

fly.io/docs/machines covers Machines API reference, machine configuration options (image, env, services, resources), GPU machine guide, volume management, private networking, and exec usage. Getting started: install fly CLI (curl -L https://fly.io/install.sh | sh), fly auth login, fly apps create, then Machines API calls. CLI wraps Machines API for common patterns. Raw API usage documented with cURL examples. Community via Fly.io Community Forum and Discord. Fly.io team publishes detailed technical blog posts explaining architecture. Strong developer community with many production deployments.