Blog/Engineering

Engineering

Running AI agents inside your own cloud account

AnjiDevOps Engineer, EnWithAI

3 min read

Talk to us

What stays inside your boundary

  1. CodeRead from your repositories, stored in your account
  2. SpecsWritten, locked and kept in your account
  3. RecordingsCaptured and masked as agreed
  4. ModelAn endpoint you choose, agreed in writing

Where the model runs is the one decision that moves data out. Make it on purpose.

The first question most security teams ask about AI agents is not about the model. It is: where does our code go? For many of the teams we work with, the only acceptable answer is “nowhere you have not agreed to”.

The short answer

Running AI agents in your own cloud account keeps code, specs and data inside your boundary. The main decisions are which model endpoint to use, how secrets and repository access are scoped, which network egress is allowed, and how the platform is upgraded.

Why run it yourself

A hosted AI tool is quick to try. But an engineering platform reads your source code, your tickets, your design documents and sometimes recordings of your systems. For regulated teams, that alone rules out sending it to someone else’s infrastructure by default.

Deploying in your own cloud account or data center flips the default. Everything is stored inside your boundary, and anything that leaves it does so because you decided it should.

The model endpoint is the real decision

The platform can run entirely inside your account. The model may not. This is the decision that determines whether code leaves your environment, so it should be explicit and written down.

Model endpoint optionsData leaves your account?Trade-off
Model in your own cloud accountNoStrongest boundary; limited to the models your provider offers there
Self-hosted open modelNoFull control; you run and scale the GPUs, and quality varies by task
Approved external providerYes, under contractWidest model choice; needs your provider review and data terms

Pick the model endpoint before the first repository is connected, not after.

Scope access like any other service

  • Read-only first. Start with read access to repositories and a test environment. Write access, such as opening pull requests, is a separate sign-off.
  • Secrets in your vault. Repository tokens and model keys live in your own secrets manager, rotated on your schedule.
  • Least-privilege identities. Each component gets its own identity with only the permissions it needs.

Control egress

Agents are good at reaching for the internet: a package registry, a documentation site, an API. Decide what is allowed. We default to denying outbound traffic except to the model endpoint, your source control and the registries your builds already use, and we log everything else that is attempted.

Upgrades without surprises

Running it yourself means owning upgrades. We ship versioned releases that your team applies through your own change process, with release notes that say what changed in behavior, not only in code. A staging deployment in the same account catches problems before production does.

What your security team will ask

Expect questions on data flows, identities, logging, model terms and incident response. Have a data-flow diagram that shows every component and every connection that crosses your boundary. It answers half the questionnaire on its own.

In CrewPE

CrewPE is deployed in your own cloud account or on your own servers, starts read-only, and uses a model endpoint agreed in writing before work starts.

See how CrewPE is deployed

Questions teams ask

Can an AI engineering platform run entirely inside our own cloud account?
The platform itself can, with code, specs and data stored inside your boundary. Whether code reaches a model outside your account depends on the model endpoint you choose.
Which model endpoint should we choose?
A model hosted in your own cloud account gives the strongest boundary. A self-hosted open model gives full control at the cost of running it. An external provider gives the widest choice and needs a provider review and data terms.
What access does a self-hosted AI agent platform need?
Read access to repositories and a test environment to start. Write access, such as opening pull requests, should be a separate, explicit sign-off, with secrets kept in your own vault.

Anji

DevOps engineer at EnWithAI, building and running CrewPE deployments in customer cloud accounts and on-premises.

Keep reading

Bring one real change. See how it would be specified, built and proven.

Start with a 45-minute call about your system: how it is built, how work is agreed today, and a sensible first step.