Northflank vs Fly.io Sprites: which platform fits your requirements?

Northflank and Fly Sprites both provide isolated environments for running AI agents and untrusted code, but they take different approaches. Fly Sprites package Firecracker microVMs as persistent Linux environments with state that survives between sessions, while Northflank provides production-grade ****microVM-backed Sandboxes alongside the broader application stack, including APIs, workers, jobs, managed databases, GPU workloads, preview environments, and BYOC into your own cloud.

The key difference is scope. Sprites are built around persistent agent environments, while Northflank combines isolated execution with the infrastructure needed to build and run production applications around it. This comparison covers lifecycle, isolation, persistence, networking, GPU support, developer experience, security, pricing, and when each platform fits.

Product and plan details are current as of September 2026.

Fly Sprites are persistent by default. Their filesystem survives pauses, so installed packages, repositories, databases, and other files remain available when the Sprite wakes. Sprites automatically move between active, warm, and cold states, with warm wakes preserving process state and cold wakes starting processes again.

Northflank Sandboxes use an ephemeral root filesystem by default. Teams can attach persistent volumes when repositories, datasets, models, or agent state need to survive restarts and scale-to-zero. Scaling a Sandbox to zero stops compute billing while keeping its volume and configuration intact.

The result is a different persistence model. Sprites make persistence automatic, while Northflank lets teams choose what should persist and what should remain ephemeral. That is useful when the same sandbox infrastructure needs to support both short-lived code execution and longer-running workloads.

Fly Sprites can boot in under one second from a cold state, while warm wakes can return in 100–500 ms with process state preserved. Checkpoints capture the filesystem state using copy-on-write, allowing developers to create reusable states without interrupting the running Sprite. Sprites currently use a fixed 8 vCPU configuration, while memory is managed automatically by the platform.

Sprites do not currently autoscale, so scaling out means provisioning additional Sprites rather than resizing an existing one. Automatic pause and wake handles idle workloads, while the Tasks API can keep a Sprite active while work is running.

Northflank Sandboxes boot in under one second and support configurable CPU and memory resources. Templates let teams define and reproduce platform resources and workflows rather than relying on VM snapshots, and Northflank has demonstrated 100,000+ concurrent Sandboxes in its 2026 Scale Invitational benchmark by ComputeSDK.

Northflank Sandboxes support configurable CPU and memory resources and scale-to-zero. At the broader platform level, Northflank also supports horizontal autoscaling based on CPU, memory, requests per second, and custom Prometheus metrics.

Each Sprite includes up to 100 GB of persistent storage, with data continuously synced to durable storage. Sprites also provide a unique HTTPS URL and configurable network policies for controlling outbound access. Fly's Connectors broker credentials for services such as GitHub, Anthropic, OpenAI, and S3-compatible storage, so Sprites can use external APIs without directly holding the underlying credentials.

Northflank Sandboxes use an ephemeral root filesystem with optional persistent volumes. Networking includes public and private ports, private service discovery, multi-project networking, network policies, static egress, and BYOC VPC routing. Secret groups provide encrypted credential injection for services and Sandboxes.

Northflank also supports GPU-backed Sandboxes, model-serving services, and jobs on supported GPU infrastructure. This means teams can combine isolated execution with GPU inference and other accelerator workloads without moving that compute to a separate platform.

Fly Sprites centers the experience on the sprite CLI and API. Developers can create Sprites, run commands, open interactive sessions, manage services, and create checkpoints from the CLI, while Sprites provide common development runtimes and AI tools out of the box.

Sprites also provide agent-focused integrations through MCP, SDKs, and Connectors, making it straightforward to give coding agents an isolated environment with access to external services.

Northflank provides Git-to-deploy, Dockerfiles, Buildpacks, external images, Workflows, Templates, GitOps, API, CLI, and SDKs.

For AI development, Northflank Skills let agents such as Claude Code, Codex, Cursor, and OpenCode deploy and manage infrastructure, while Harnesses provide dedicated cloud environments for AI coding agents with their own resources and networking.

Sprites are primarily focused on isolated execution environments. They provide persistent compute, services, URLs, and agent tooling, but the broader production architecture around an application, such as APIs, databases, and other services, needs to be built separately on the Fly platform or elsewhere.

Northflank Preview Blueprints create complete preview environments with services, databases, jobs, volumes, networking, and configuration. The same platform provides managed databases, GPUs, queues, persistent infrastructure, and the deployment workflows needed to move that architecture into production.

Northflank also supports Northflank Cloud, BYOC, and BYOK, so teams can run the sandbox and its supporting application infrastructure on Northflank's managed cloud, in their own cloud account, or on eligible existing Kubernetes infrastructure.

Each Sprite runs in its own Firecracker microVM with its own kernel, filesystem, and network namespace. Sprites also provide outbound network policies and Connectors that broker credentials for external services without exposing the underlying provider token to the Sprite.

Sprites currently do not provide a fine-grained RBAC model for controlling access to individual Sprite resources. Authentication uses Fly.io accounts and scoped API tokens, while agent access can be restricted through Connector policies.

Northflank Sandboxes use microVM-backed isolation with a separate kernel, while the platform adds granular RBAC, SAML/OIDC SSO, directory sync, scoped API access, and audit logs. Northflank Enterprise also supports forward-deployed and air-gapped environments for stricter infrastructure requirements.

Northflank also provides logs, metrics, health checks, alerts, and log sinks across sandboxes and the rest of the application stack. This gives platform teams security and observability controls at both the workload and platform level.

Fly Sprites bills CPU and memory only while a Sprite is active, while persistent storage continues to incur storage charges. Current published rates are $0.07 per CPU-hour and $0.04375 per GB-hour of memory, with storage priced separately. Sprites also offer subscription plans with included usage and concurrency limits.

Northflank Cloud charges $0.01667 per vCPU-hour and $0.00833 per GB-hour of memory, billed per second with no seat fees on Pay-as-you-go. Sandboxes use the same resource-based model, while GPU and storage resources are billed separately.

At the published rates, Northflank's CPU and memory prices are substantially lower than Sprites'. Northflank also lets teams configure sandbox resources to match the workload rather than using Sprites' fixed 8-vCPU configuration. For teams that need GPUs, supporting application infrastructure, or BYOC, Northflank can also consolidate more of the stack under the same platform.

Fly Sprites fit teams that want persistent Linux environments designed around AI agents and arbitrary code execution. Sprites keep filesystem state across pauses and provide Firecracker isolation, checkpoints, Connectors, and agent-focused tooling.

Northflank fits teams that need production-grade isolated execution as part of a broader platform. Sandboxes can run alongside GPUs, volumes, APIs, workers, jobs, managed databases, previews, and workflows, while Northflank also supports Northflank Cloud, BYOC, and BYOK for customer-controlled infrastructure.

Fly Sprites are persistent Linux environments built around AI agent and code execution, while Northflank Sandboxes provide isolated execution as part of a broader application and AI infrastructure platform.

Yes. Each Sprite runs in a Firecracker microVM with its own isolated environment.

The Sandbox root filesystem is ephemeral by default, but persistent volumes can preserve repositories, models, datasets, and agent state across restarts and scale-to-zero.

No. Sprites currently provide CPU, memory, and storage resources, without a GPU resource.

Yes. Northflank supports GPU-backed Sandboxes as well as GPU-enabled services and jobs on supported infrastructure.

No. Sprites run on Fly.io managed infrastructure. Northflank supports BYOC and BYOK for teams that need workloads on their own cloud or Kubernetes infrastructure.

Yes, for teams that need isolated agent execution alongside production services, databases, GPUs, previews, and customer-controlled infrastructure. Northflank also provides Harnesses and Agent Skills for AI coding workflows.