Sprites

Operator's take

Sprites answers one question: where should an AI agent actually run? Its own pitch is "the simplest answer for 'where should I run a blob of code'" — whether that blob is Claude Code churning through a task or a binary a user just uploaded. A Sprite is a hardware-isolated Linux computer (a dedicated microVM, not a shared container) that you create on demand, that keeps its files and installed tools between runs, and that you can snapshot and roll back. It comes from Fly.io, whose whole business is running microVMs close to users, so the "give every agent its own real computer" framing has infrastructure behind it, not just a landing page.

Be clear-eyed about who this is for out of the box: on its own, Sprites is developer infrastructure. You install a CLI, log in, create sprites, and drive them with commands, an API, or SDKs in five languages. A non-technical operator does not open the docs and immediately have a working setup. The reason it earns a place in this directory anyway is that a single operator playbook bridges the gap — "run a CLI agent in a disposable sandbox, with Slack as the front door" (see the companion playbook). Follow that one piece and a motivated non-engineer can put an AI agent to work in a throwaway box and talk to it from Slack. That's the honest boundary: the tool is developer-grade; the one bridge makes it operator-reachable, and that's exactly why it's listed rather than filed under "you'll need a developer."

The two features that make it worth the trouble, in plain terms. Checkpoint & restore: snapshot the box before you let an agent run unsupervised, and rewind in one command if it makes a mess — the thing that makes the sandbox genuinely disposable. And scale-to-zero: a Sprite pauses itself after ~30 seconds of inactivity and stops billing compute, then wakes in 100–500ms with everything where it left off — so an always-available agent doesn't mean an always-running bill. The catch worth knowing before you rely on it: your disk persists (files, installed packages, git repos, SQLite), but your processes do not survive a full cold wake — a server or agent you started by hand is gone unless something restarts it.

What it's good at

• A safe home for AI coding agents — the headline use case is running LLM-generated code (Claude Code and friends) in strong isolation, so an agent's mistakes stay inside a box you can rewind or delete, not on your own machine.
• Checkpoint & restore — snapshot the whole environment and roll back to it; the feature that turns "an agent has root on a Linux box" from scary into disposable.
• Genuinely persistent — a real 100 GB ext4 filesystem that behaves like local disk (SQLite, shared-memory files, normal Linux semantics all work); install dependencies once and they're there on every wake.
• Scale-to-zero economics — per-resource usage billing (compute charged only while active), so an idle sprite costs almost nothing; a ~4-hour Claude Code session runs on the order of $0.44.
• Credential-safe connectors — wire Slack, GitHub, OpenRouter, or any HTTP API once at the org level; the Sprite calls a gateway URL and never holds the token, with deny-by-default access policies per sprite.
• Meets you where you code — CLI, REST API, and SDKs for JavaScript, Go, Elixir, and Python; a remote MCP server to manage sprites from an MCP client; and a Managed-Agents integration to run Anthropic agent sessions in a sandbox you own.

What it's not

• Not no-code, and not turnkey for a non-technical operator — you work through a CLI, API, and tokens. It's listed here only because one bridge playbook makes it operator-reachable; without that piece, this is developer infrastructure.
• Not a general web-app host — it's a sandbox for running code and agents, not a Vercel/Netlify-style place to deploy a site. (Its parent Fly.io is the tool for that.)
• Not a persistent always-on server — it's designed to pause when idle; processes you start by hand don't survive a cold wake, so anything that must stay up needs Sprites' services/tasks mechanism, not a bare nohup.
• Not free — there's no free tier in the plan lineup; paid subscriptions start around $20/month plus usage. Cheap for what it is, but a cost from day one. [ambiguous — pricing from secondary sources; verify on the live pricing page]
• Not battle-tested-old — it launched in January 2026. The Fly.io microVM foundation is proven; Sprites as a product is new.