AI DevOps Agent · Open Source

The DevOps agent
that ships the
whole pipeline.

Give Pilot a GitHub repo and a cloud provider — it writes the Dockerfile, Terraform, Helm chart, and CI/CD pipeline, then monitors and scales your service automatically.

pilot — deploy
< 6min
git push → live service
full pipeline, zero manual steps
~72%
avg compute cost cut
KEDA scale-to-zero + Karpenter
10
pipeline steps automated
Dockerfile → deploy → monitor
6
languages auto-detected
Node · Python · Go · Java · Rust · Ruby
3
clouds supported
AWS · Azure · GCP
0
secrets in generated output
OIDC + ESO, no static keys
The state of DevOps

Your infra is duct‑tape, your
on‑call rotation is resentment,
and every deploy is a séance.

01 — TOIL

3 a.m. pages for the same flaky pod

Your strongest engineer is awake at 3:14, copy‑pasting kubectl describe into Slack for the fourth night this week.

02 — DRIFT

“It works on my staging”

Six environments. Four Terraform repos. One Notion doc that was last accurate in March. The config has drifted, and so has the team.

03 — COST

AWS bill nobody can defend

Three unused RDS instances. A NAT gateway from a 2022 spike. Idle GPUs in eu‑west‑3. Finance keeps asking. Nobody knows.

04 — INCIDENT

Rollbacks are folklore

You thinkyou can roll back. You're not sure who has the latest good SHA. The runbook lives in a Notion doc that 404s on mobile.

// the platform

One agent. The entire path from git pushto “still healthy at 4 a.m.”

Pilot owns infrastructure as a first‑class actor. It writes the IaC, picks the cloud, runs the deploy, watches the metrics, fixes what it can, and pages a human only when it really, genuinely, should.

/ service graph
api · okworker · scalingpostgres · primarycdn · 99.99queue

Live service graph, owned

Pilot maps every service, queue, db, and edge node it ships — and keeps the map alive as the topology changes.

/ pipelines
gitpushbuild✓ 1m12stest✓ 412 okcanary3% liveprod14:0214:0314:0414:0514:06

CI/CD that reasons

Canary, blue/green, or progressive rollout — Pilot picks the strategy and bails out the moment the SLO budget twitches.

/ chat ops
$pilot, why is checkout slow in eu-west?
↪ scanning traces · 14s window
73% of latency from payments-svc cold starts
bumped min-replicas 2→4 in eu-west-1
p99 back under 220ms

Talk to your infra

Plain English in. Diffs, deploys, and incident commentary out.

/ cost

Cost as a first‑class metric

Pilot right‑sizes, schedules, and kills the orphaned NAT gateway you forgot.

/ rollback

Self‑healing rollouts

Detect bad release, halt rollout, revert. No paging, no Slack thread.

// connections

Plays nicely with everything already in your stack.

No rip‑and‑replace. Pilot reads your existing Terraform, your Helm charts, your GitHub Actions, and your runbooks — and just gets to work.

AWS
GCP
Azure
k8s
Docker
tfTerraform
GitHub
GitLab
Datadog
PagerDuty
Slack
Prometheus
Grafana
SQLPostgres
Redis
Vercel

+ 80 more · open MCP & REST APIs · BYO tools via shell wrapper

// how it works

Three boxes. Connect them. Walk away.

STEP · 01

Point Pilot at your repo.

One pilot init. It reads your code, your existing IaC, your env. No magic — it tells you exactly what it plans to take over.

repo→ pilot
STEP · 02

Approve the plan.

Pilot drafts the IaC, the pipeline, the dashboards, the alerts. You diff it, comment, merge. Then it ships, end to end.

PLAN+ 12 resources~ 3 modified- 0 destroyedapprove →
STEP · 03

It runs the night shift.

Pilot deploys, watches, scales, patches, rotates secrets, and writes the postmortem. You sleep through the on‑call rotation you used to dread.

zzZ

// ALL THREE STEPS, FIRST DEPLOY, UNDER 11 MINUTES MEDIAN

Read the technical brief
// slash commands

Every tool, one
slash away.

16 Claude Code commands that run directly in your editor. Infra ops without leaving your terminal.

/deployFull pipeline — Dockerfile → Terraform → Helm → CI/CD
/terraformGenerate Terraform for AWS · Azure · GCP
/sre-guardStart the monitoring daemon — watches all services
/optimize-costApply KEDA scale-to-zero + Karpenter to your cluster
/auditSecurity + drift audit on existing infrastructure
/helmGenerate + lint a production Helm chart
claude code
✓  payment-api repo loaded · aws · us-east-1
// security & trust

A robot in production needs guardrails, not just guts.

Pilot operates with least‑privilege roles, dry‑runs every change, and asks for a human signature on anything that touches money, identity, or data.

/ guardrails

Production‑grade by default

Least privilege, audit log on every action, dry‑runs on every change, human approval on every blast radius.

/ approvals

Human in the loop, when it matters

Approvals routed by policy. Pilot waits patiently. Everything else, it just ships.

/ audit log
14:02:11 deploy api v126→v127
14:04:48 scaled worker 4→6
14:09:02 ! rotated kms-key pay-2
14:11:33 closed alert cpu-eu1
14:18:09 applied tf plan #3a91
14:24:17 ! blocked: prod ddl needs sign-off
14:31:50 patched cve-2026-19284
14:42:09 nightly snapshot · 12 dbs
15:01:00 rebalanced traffic eu→us 8%

Every action, signed and queryable

Append‑only audit log, exportable to your SIEM. Ask Pilot anything about its own behaviour.

/ secrets

Secret rotation, hands‑off

Pilot rotates, distributes and revokes — auto, on schedule, on compromise.

/ self‑hosted

Runs in your VPC

Your cloud, your account, your data. Pilot's control plane never touches it.

/ policy
# policy.rego — pilot allow/deny
deny {
  input.action == "drop_table"
  input.env    == "prod"
}
allow {
  input.action == "scale"
  input.replicas <= 32
}
↪ enforced on 142 actions today · 0 violations

Policy as code

OPA‑native. Encode “Pilot may never…” once, enforce it everywhere.

// faq

The questions everyone asks first.

Hover any question to peek the answer. Click to keep it pinned open.

Is Pilot just a wrapper around an LLM?

No. Pilot is a long‑running agent with persistent memory of your infra, a typed tool layer over cloud APIs, deterministic planners for IaC diffs, and a model‑agnostic reasoning core. The LLM is one of several substrates — swapped per task. Cost, latency and determinism win.

What clouds and stacks does it support?

AWS, GCP and Azure on day one. Kubernetes (EKS / GKE / AKS / vanilla), Nomad, Fly.io, bare‑metal via SSH. IaC: Terraform, Pulumi, CDK, Helm. CI: GitHub Actions, GitLab, CircleCI, Buildkite. Observability: Datadog, Grafana, Honeycomb, OpenTelemetry. If it has an API or a shell, Pilot can drive it.

Can Pilot really push to production unattended?

Yes — within the guardrails you set in policy. By default, anything inside the staging blast radius is fully autonomous. Production deploys require either passing canary metrics or a human approval, depending on your policy.rego. Most teams ramp Pilot's autonomy week by week.

What happens when it gets something wrong?

Pilot only takes actions it can reverse. Every change is preceded by a dry‑run, the diff is recorded, and the inverse action is queued before execution. When SLOs degrade post‑deploy, Pilot rolls itself back automatically and pages the on‑call with a complete incident packet (root cause, blast radius, mitigation).

Do you train on our code or our infra metadata?

Never. Your code, your secrets, and your infra topology never leave your boundary in the self‑hosted plan, and are not used to train any shared model in the managed plan. Logs and traces are encrypted at rest with your KMS keys.

How is this different from Copilot / Cursor / Devin?

Those agents help engineers write code. Pilot replaces the engineer that operates code — pipelines, infra, observability, incident response, cost. It's not a sidekick in your editor; it's a teammate in your on‑call rotation.

How much does it cost?

$0 to start: 1 environment, unlimited deploys. Team plan from $40 / engineer / month. Self‑hosted enterprise pricing is flat by fleet size. Most teams report Pilot pays for itself in cloud savings inside 60 days.
// ready when you are

Give your DevOps to a machine.
Get your weekends back.

Eleven minutes from pilot init to your first autonomous deploy. No card. No sales call. No yak shaving.