Your coding agents, always on

Run Claude Code, Codex, OpenCode, Pi, or Cursor in the cloud, each in its own microVM with a persistent workspace and your repositories. Start a task, disconnect, and pick it up from any device.

  • Codex
  • Claude Code
  • OpenCode
  • Pi
  • Five agentsClaude Code, Codex, OpenCode, Pi, Cursor
  • Kata microVMOne per harness
  • Live previewsPublic URL per port
  • Northflank Cloud or BYOCRuns in your VPC

The runtime behind 2,000+ ambitious teams

Customer stories
  • govtech logo
  • chaidiscovery logo
  • figma logo
  • sentry logo
  • coreweave logo
  • upwork logo
  • weightsandbiases logo
  • infisical logo
  • isembard logo
  • edb logo
  • rothys logo
  • v7labs logo
  • directus logo
  • zamp logo
  • projectdiscovery logo
  • factorio logo
  • smallestai logo
  • scaler logo

What you can run

Your existing coding agent, on dedicated compute that keeps running when you disconnect.

The agent you already use

Claude Code, Codex, OpenCode, Pi, and Cursor, signed in with an API key or your account

Quickstart

Hand it a task, come back to a branch

Runs independently of your machine, and pushes to its own branch

Quickstart

From prompt to preview to production

Preview each change on a public URL, then release through your pipelines

Explore preview environments

Our cloud or yours

Run harnesses on Northflank Cloud, or on a BYOC cluster in your VPC

Explore bring your own cloud

Long-running tasks

Harnesses run independently of your machine. Start a task, disconnect, and review the branch and preview later.

Runs while you’re disconnected
Closing your laptop or terminal doesn’t stop the agent.
Branch-scoped pushes
Git credentials are scoped to the harness branch and refreshed automatically.
Preview URLs
Expose the app on a public URL and review it from any device.

The software lifecycle, on a loop

Agents pick up tasks, write code, and open pull requests. Northflank turns every change into a live preview, promotes it through staging to production, and feeds what happens in production back into the next task. No tickets, no handoffs, no waiting for environments.

  1. Cloud agentHarness in a microVM
  2. Code generationYour repo, in a sandbox
  3. PreviewLive environment per PR
  4. StagingPromoted through your pipeline
  5. ProductionRelease flows and rollbacks
  6. OperationsLogs, metrics, alerts, backups
  7. ↺ Production tells agents what to fix next
The old loop
Ticket queueEnvironment requestReview backlogRelease window
On Northflank
Every step automated, nothing waiting in between

Access

Connect to the same workspace from the dashboard, the CLI, or another machine.

Browser terminal
Multiple terminal windows in the Northflank dashboard.
SSH and ports
SSH in and forward ports with the Northflank CLI.
Pause and resume
Pause to stop compute. Resume with the workspace unchanged.

API access from the harness

Agents can create sandboxes, deploy services, and provision databases through the Northflank API and CLI.

Quickstart
terminal
npm install --global @northflank/cli
northflank login

northflank dev ssh --projectId my-app --harnessId claude-code
claude --continue
Output
  1. ✓ Connected to claude-code
  2. harness@claude-code:~/my-app$ claude --continue

Team controls

Policies for which agents run, and who can access each harness.

Allowed agents
Restrict which coding agents a team can run.
Private harnesses
Harnesses whose workspace only their creator can access.
Projects, RBAC, and secrets
The same access controls as the rest of Northflank.

Harness features

A Linux environment with your code, your tools, and the compute to run them.

Quickstart
  • Supported agents

    Claude Code, Codex, OpenCode, Pi, and Cursor, or an empty harness for your own tooling.

  • A microVM per harness

    Each harness runs in its own Kata microVM, isolated from other workloads.

  • Persistent workspace

    The home directory is a 5 to 100 GB volume that persists across restarts.

  • Repositories

    Clone a branch or create a new one, plus up to 10 more repos. Pushes are scoped to the harness branch.

  • Live previews

    Expose ports on public URLs to preview what the agent is building.

  • Compute

    Any plan size and an optional GPU, with environment variables and secret files.

Pricing

Standard Northflank rates, billed per second. No seat or per-agent fees. Paused harnesses don’t use compute.

Full pricing
Billing period
Resource
Billing period
Northflank CloudContainers, by the secondNorthflank BYOC (Your VPC)Managed node vCPU only, by the second
Compute
CPU$0.01667/ vCPU / hour$0.02083/ node vCPU / hourVolume discounts available under a CSA
Memory$0.00833/ GB / hourNot billed
Ephemeral storageNot persistedFreeNot billed
Storage and network
Persistent diskNVMe, for addon and service volumes$0.11657/ GiB / monthNot billed
Additional IOPSFirst 3,000 IOPS and 140 MiB/s per disk included$7.37/ 1,000 IOPS / monthNot billed
Additional throughputFirst 3,000 IOPS and 140 MiB/s per disk included$5.85/ 100 MiB/s / monthNot billed
Network egressIngress is free$0.06/ GBNot billed
Egress gatewayDedicated IP$15.00/ gateway / monthNot billed

Customers

Teams that build and ship with Northflank, in their own words.

  • PolyguardShipping every change, fast

    Cycle time is everything. With Northflank, I can make 100 commits and 100 deployments in a single day – it keeps up with my pace like nothing else. When it comes to debugging, the meantime to resolution is unbeatable. I can identify issues and deploy fixes faster than customers can even report them. I know about problems before they do and have them fixed before they call.

    Joshua McKenty – profile image
    CEO @ Polyguard
  • PebblebedA whole product, end to end

    Northflank is the first batteries-included developer platform that doesn’t suffer from the invisible ceilings that hover over its competitors. We could have built all of Slack with Northflank – and we would have, had it been available.

    Keith Adams – profile image
    Formerly Chief Architect Slack, Facebook, VMware, GP @ Pebblebed
  • WeightsGPU training and inference

    Northflank helped us scale to over 3M users with an engineering team of two. It allows us to run 10,000 AI training jobs and half a million inference runs per day, without us having to worry about autoscaling or spot instance orchestration. We don’t waste time or money on infrastructure, so we can focus on building product.

    JonLuca DeCaro – profile image
    Founder @ Weights, acquired by OpenAI

Ready for enterprise

Enterprise

Run agents, AI-built software, and core systems inside your own infrastructure, with the compliance, governance, and guarantees your organisation needs.

Deploy now
  • Bring your own cloudFully managed clusters in your own VPC, on AWS, Google Cloud, Azure, Oracle, or bare metal.
  • Forward-deployed control planeRun the control plane in your own infrastructure, air-gapped if you need it. No platform data leaves your network.
  • GovernanceSSO and SAML, SCIM and JIT provisioning, RBAC, and audit logs for every team and agent.
  • Security and isolationMicroVM sandboxing, WAF and DDoS protection, and workload identities.
  • 99.99% SLA and supportAn uptime SLA, and a dedicated Slack channel straight to our engineers.
  • Volume discountsUsage discounts on Northflank Cloud and BYOC as your spend grows.

Questions

Contact us
Which coding agents can I run?

Claude Code, OpenAI Codex, OpenCode, Pi, and Cursor. You can also start an empty harness and install your own tools.

How do agents sign in?

Claude Code, Codex, and Cursor can use an API key or your account. OpenCode and Pi sign in with your account from the terminal. API keys are stored encrypted and never shown again.

What happens to my files when a harness restarts?

The workspace is on a persistent volume, from 5 to 100 GB, so files in the home directory survive restarts and pauses. You can grow the volume later.

Can the agent push to my repositories?

Yes. Harnesses push through a proxy scoped to the harness’s branch, with credentials that refresh automatically.

Can I run harnesses in my own cloud?

Yes, on BYOC clusters with microVM (Kata) support, so code and data stay in your account.

How are harnesses billed?

Like any workload: by the second, for the CPU, memory, and GPU of the plan you choose, plus the workspace volume. On the free Sandbox plan, a harness can use up to 1 vCPU and 1 GB.

Run your first harness

Start free on Northflank Cloud or in your own cloud account. Same platform and workflow either way.

npm i -g @northflank/clinorthflank login