InfrapilotAlpha

Autonomous cloud infrastructure

Pilot is an agent that runs your AWS, Google Cloud or Azure account. You write a few lines of intent: what must stay fast, what it may cost, what it must not touch. Pilot sizes, scales and repairs to keep that true. Every change is checked by code against your limits and recorded with its reason. If a health check fails afterwards, the change is undone.

In alpha: free, and rough in places. What that means

Ledger

Example: one day of fleet shop on AWS

Pilot is running

checkout-api p99

204 ms must stay below 400

Monthly cost

$68 ceiling $150

Pilot

Scaled checkout-api down as traffic fell

MemorySize1024 to 512 MB

Comfortably under the objective for two hours. Saves $41 this month.

Pilot

Undid the console change

MemorySize256 to 1024 MB

It broke the latency objective.

Outside

Someone changed checkout-api in the console

MemorySize1024 to 256 MB

Made outside Infrapilot, by sam. p99 went to 455 ms.

dana

Approved. Pilot may now use up to 8 workers

bounds2 to 5 to 2 to 8

Pilot asks

Raise the worker concurrency bound to 8

The orders backlog is 4,000 and five workers cannot clear it. If the answer is no, orders wait about nine minutes.

Rules

Set worker concurrency to 12refused

Outside the bound of 2 to 5 set in the spec.

Pilot

Gave checkout-api more memory

MemorySize512 to 1024 MB

Traffic doubled. p99 reached 612 ms against an objective of 400.

Engine

Health check passed on checkout-api

Pilot

Built checkout-api, a queue and a bucket

resources0 to 3

Three workloads in infrapilot.yaml did not exist yet.

Pilot

Reviewed the fleet for the first time

Nothing to adopt. Three workloads to build, inside a ceiling of $150 a month.

dana

Saved the fleet spec: one intent, three workloads

Chain intact: 96 entries, every hash follows

The ledger above is an illustration. Watch a recorded run on a simulated account, entry by entry, without signing in.

Two ways in

Both end in the same place: a short spec in your repository, and Pilot keeping it true in your own cloud account.

Adopt

Pilot runs the cloud account you already have

Tag the resources you want handled. Pilot adopts them as they are, works out which settings your autoscalers already own, and tunes the rest toward your objectives: memory, concurrency, task counts, capacity limits. Your deploys and your CI stay as they are.

For teams on ECS, Lambda, Cloud Run, Container Apps or App Service, with a managed database and nobody whose job is infrastructure.

How setup works

Build

Leave Render, Railway or Heroku for your own cloud account, with no infrastructure code

Describe the app: a web service, a worker, a Postgres database, a Redis, a cron job. Pilot builds it in your own AWS, Google Cloud or Azure account. Then it stays: the same agent keeps it sized, scaled and healthy after the move.

For teams outgrowing a platform's limits or its bill. What Pilot builds are ordinary resources in your account, with nothing proprietary to migrate off later.

The move, step by step

You choose how much Pilot may do

Each intent in the spec has a level. Autonomous is the default; the other two are for the first week, or for the parts you never want changed without you.

Observe
Pilot reads, reviews and explains: what runs, what it costs, where an objective is at risk. It changes nothing.You grant a read-only role.
Propose
Pilot prepares each change with its evidence and its cost. The change waits for one click from you in the dashboard, and runs only if nothing has moved since.You grant the tune role.
Autonomous default
Pilot acts inside the bounds you set and asks a person when it reaches one.You grant the tune role, or the build role if Pilot creates resources.

At every level the same engine checks each change against your bounds and cost ceiling, health-checks it, and undoes it if the check fails.

How it works

  1. 01

    Connect your cloud

    Paste one prompt into your coding agent. It connects your cloud account in minutes.

    Prompt
    Set up Infrapilot for this project. Read
    https://infrapilot.fly/docs/onboard.md
    and follow it from start to finish.
    
    Setup token: shown when you create a workspace
  2. 02

    Say what you need

    A few lines of intent, not infrastructure code: what to run, how fast it must be, what it may cost.

    infrapilot.yaml
    intents:
      - name: checkout API stays fast
        objectives:
          - {resource: api, metric: Duration,
             stat: p99, below: 400}
        max_monthly_cost_usd: 150
  3. 03

    It runs itself

    Infrapilot builds, scales, tunes and repairs around the clock, and asks you only when it reaches a limit you set.

What it replaces

Infrastructure code for each new service

A few lines of intent

Capacity planning

Sized from live metrics, around the clock

Finding out from a customer

Acted on while an objective is still only at risk, with the reason written down

Surprise bills

A cost ceiling Pilot's changes cannot cross

You stay in control

Read the security model, or send your security reviewer the one-page overview (PDF).

What alpha means

Infrapilot is in alpha: the first period in which other companies run it on their own accounts. It is free and by invitation. If anything goes wrong in your workspace, you hear from the founder the same day.

Pricing

The alpha is free. These are the plans that follow it; each will start with 14 days of Team, free, with no card.

Starter

$199 a month

One product on one cloud account.

  • 1 environment on AWS, Azure or Google Cloud
  • 25 managed resources
  • 750 agent decisions a month
  • Requests to you by email
  • A full review every 6 hours
Request access

Team

$599 a month

Production and staging, no DevOps team needed.

  • 3 environments, any mix of clouds
  • 150 managed resources
  • 3,000 agent decisions a month
  • Requests by email, Slack, Teams or phone
  • A full review every 4 hours
  • Ledger and state export
Request access

Scale

$2,400 a month

Many services across many accounts.

  • 10 environments
  • 1,000 managed resources
  • 15,000 agent decisions a month
  • Requests by email, Slack, Teams or phone
  • A full review every 2 hours
  • Ledger and state export
Request access

Enterprise

Custom

Regulated or very large estates.

  • Dedicated control plane
  • Agent sandboxes inside your VPC
  • Your own model provider (Bedrock, Vertex AI, Foundry)
  • Custom rules and approval chains
  • Uptime SLA and a named engineer
Talk to us

Nothing is charged during the alpha. Use beyond a plan's allowance is shown in the dashboard and is not charged either.

Questions

Who is this for?

Teams of roughly 3 to 30 engineers with nobody whose job is infrastructure: either on AWS, Google Cloud or Azure with containers or functions and a database, or on Render, Railway, Heroku or Fly and outgrowing it. If you are happy on Vercel and Supabase with nothing else, you do not need this. If you run mostly on Kubernetes, not yet.

Is it safe to let an AI change production?

It can only make changes your limits allow, and an engine it cannot bypass checks every one. Each change is health-checked and undone automatically if it fails.

What makes the decisions?

Claude, Anthropic's model, proposes each change. Ordinary code decides whether it happens: your bounds, your cost ceiling, the provider's schema, a health check afterwards. A wrong or manipulated model still cannot get past that code, or beyond the role you created.

What happens to our infrastructure code?

Keep it. Pilot works on your live resources, whichever tool created them. It detects what Terraform, CloudFormation, CDK or Pulumi manages and, by default, proposes changes to those resources instead of making them, so the two do not undo each other. For anything new you would rather not write code for, a few lines of intent take its place.

Am I locked in?

No. Your resources are ordinary resources in your own account. Remove Infrapilot's access and they keep running as they are.

We have nothing yet. Does it still work?

Yes. That is the second way in: describe the app, and Pilot builds it in your own account and then keeps running it.

Which services does it run?

Compute, queues and databases, and the front door of your app: load balancers, CDNs and API gateways. Set a goal where users feel it, like p99 latency at the load balancer, and Pilot meets it by scaling what runs behind it. See the full list for AWS, Azure and Google Cloud.

Request access

Infrapilot is in a private alpha, and the alpha is free. Leave your address and we will write to you when your workspace is ready.

Used only to reply to you. Privacy

Let your infrastructure run itself.