Nobody is on call tonight.

Setup is one prompt

Paste it into the coding agent you already use. It reads our docs, creates a scoped role in your cloud account, writes your fleet spec from what it finds in your repository, and starts Pilot. About ten minutes, most of it your cloud provisioning a role.

Set up Infrapilot for this project. Read
https://18-207-170-166.sslip.io/docs/onboard.md
and follow it from start to finish.

Setup token: your token, shown when you create a workspace

Works with Claude Code, Cursor, Codex and anything else that can read a web page and run a shell command.

How it works

1. 01

   Connect a cloud account

   Your coding agent creates one role (AWS), service principal (Azure) or service account (Google Cloud) for Pilot, scoped to what you choose. Pilot can never do more than that role allows, and the credentials exist only inside your workspace’s sandbox.

   AWS

   An IAM role that trusts Infrapilot, with an external id only your workspace has

   Azure

   A service principal on one resource group or subscription

   Google Cloud

   A service account Infrapilot may impersonate. No key leaves your project

2. 02

   Say what must stay true

   An objective is a metric that must stay under a number. Add a cost ceiling, bounds on what Pilot may set, and anything it must never touch. That is the whole configuration. There is no infrastructure code to write, review or keep in step with reality.

   intents:
     - name: checkout API stays fast
       objectives:
         - {resource: api, metric: Duration,
            stat: p99, below: 400}
       max_monthly_cost_usd: 150
       bounds:
         api.MemorySize: [128, 1024]
   boundaries:
     protected: [orders-db]

3. 03

   Pilot takes over

   It adopts the resources you tagged, or builds the ones you described. Then it watches every objective and acts before one breaks: more capacity under load, less when it is quiet, a revert when someone changes production by hand. Where your provider already runs an autoscaler, Pilot tunes it instead of fighting it.

   • api p99 latency188 msunder 400
   • orders backlog12under 300
   • events write throttles0under 1
   • monthly cost$131ceiling $150
4. 04

   It asks when it reaches a limit

   When the right move is outside its limits, Pilot does not improvise. It does what the limits allow, then sends you the exact change it proposes, with the numbers, by email, Slack, Teams or text message. You approve or deny in one tap, and the decision is recorded under your name.

   Pilot needs a decision

   Raise the events table cost ceiling from $40 to $55 a month

   Ingest doubled since Friday. Holding zero throttles needs 80 write units; the ceiling allows 61. If the answer is no, about 3% of writes keep failing at peak.

   Approve Deny
5. 05

   Read the record

   Every change, refusal and request is in a ledger with Pilot’s reasoning, each entry chained to the one before so nothing can be quietly rewritten. The dashboard shows it live beside the graph of what you run. Your resources stay ordinary resources in your own account.

   14:06

   Pilotevents autoscaler minimum 5 → 60, maximum 10 → 80

   Demand is 41 writes a second. Capacity belongs to the autoscaler, so I raised its bounds.

   entry 418 hash 9f2c41d07ab3 follows e07b6612c9d4

It cannot leave the envelope you draw

Pilot proposes. An engine it cannot bypass decides. Before a change reaches your cloud, the engine checks it against rules that are code, not instructions in a prompt:

• the provider’s own schema for that resource
• your bounds and your cost ceiling
• no more than half of any capacity removed in one change
• a value your provider’s autoscaler owns is never overwritten
• protected resources are never touched
• anything irreversible needs a named person
• a change budget that refills only while production is healthy

After a change lands, a health check runs. If it fails, the change is undone automatically and Pilot is told why.

What it takes off your plate

Your account, its own sandbox

Each workspace’s agent runs in its own locked-down container: a read-only filesystem, no inbound network, your cloud role and your ledger and nothing of anyone else’s. Our model key never enters it, and neither does any other customer’s data.

Pricing

Priced on what Pilot manages and how much it has to think. Every plan starts with 14 days of Team, free, with no card.

A managed resource is anything Pilot watches or changes. A decision is one time Pilot wakes up and reasons about your fleet; a quiet fleet uses a handful a day. Beyond your plan it is $4 per resource and $0.20 per decision, so a spike never stops the agent. Pay yearly and two months are free. Team costs less than 5% of one infrastructure hire.

Questions people ask first