Security

Security you can verify

The code is open, the isolation is documented and you decide where everything runs. Here is what the platform does, and what it does not do yet.

What your IT checks

An honest status for each point: available, in progress, on our roadmap or not yet.

  • Hosting on your premises, down to an isolated environment

    On your servers, with a local model and your images.

    Available
  • Data sent to models

    Depends on the model you choose: a local model stays with you.

    Available
  • Encrypted secrets, never seen by agents

    AES-256-GCM encryption, credential proxy.

    Available
  • Isolation of every run

    One network and containers per run, with Docker.

    Available
  • Run history

    Who, when, which model, what cost, which tool calls.

    Available
  • Viewable audit log

    Permission changes are recorded in the database; viewing them in the interface is on our roadmap.

    On our roadmap
  • Team sign-in

    Email and password, magic link, Google or GitHub.

    Available
  • Data processing agreement and subprocessors

    Being written.

    In progress
  • Enterprise SSO (SAML, Microsoft Entra)

    With the Enterprise license, per organization.

    Coming soon
  • SOC 2 and ISO 27001

    Not started. The code is open for audit.

    Not yet

See the Security page

Access control

Who can do what, and with which key.

Permissions per space

Five ready-made roles (admin, builder, operator, runner, viewer) and custom roles. Each role opens precise permissions: agents, runs, files, connections.

Capped API keys

An API key belongs to a space and can never do more than the person who created it.

Team sign-in

Email and password, magic link, Google or GitHub. You can close sign-ups or limit them to your domains.

Users of your applications

Appstrate can act as an OpenID Connect provider for the users of your own applications, with OAuth clients per space.

Agent isolation

What happens when an agent runs.

One network per run

With Docker, every run gets its own internal network and an unprivileged container. The agent has no route to your host, to other runs or to the Internet, except through the sidecar.

Agents never see your secrets

The agent calls a tool, and a platform proxy adds the credential at call time. Neither token nor model key enters the agent's container.

Controlled outbound traffic

Each integration declares the addresses it may call. Redirects are checked, and private ranges and cloud metadata are blocked.

Encrypted secrets

Credentials and keys are encrypted at rest with AES-256-GCM, with a versioned key and a rotation script.

Limits per organization

By default, at most 200 launches per minute and 50 concurrent runs, with caps on duration, memory and CPU per run.

Docker in production

Docker mode is the production install. The mode without isolation is for development and evaluation only.

Traceability

What the platform records.

Run history

Every run keeps who started it, the model, the result, the cost and the tool calls.

Audit log

Changes to permissions, members, keys and connections are recorded in the database, append-only, with their author. Viewing them in the interface and through the API is on our roadmap.

Calls carrying a credential

Every call that could send a credential is recorded: who, which connection, which host, which result.

Sovereignty

Where Appstrate runs, and what it sends.

On your premises

Install on your servers, with PostgreSQL and files on disk or on S3 storage. Your data stays with you, except what you send to an external model or tool.

Without Internet access

The install can run with no Internet access: images in your registry, a local model, internal storage and model catalog reading turned off. The SaaS tools you plug in, however, need outbound Internet access.

No usage telemetry

Appstrate sends no usage data. The only outbound request by default is reading the model catalog, which you can turn off.

Open code

Apache 2.0, with no usage limits. Your team or an auditor can read it all.

Report a vulnerability

Write to [email protected] or open a security advisory on GitHub. We acknowledge within 48 hours and assess within 5 business days. Fix targets: 7 days for a critical flaw, 14 days for high severity, 30 days for medium severity and 90 days for low severity.

Let's see what it would look like on your side.

Try it on your own in a few minutes, or take thirty minutes with us to scope a first use case.

To try it

Try it for free

Create a cloud account and describe your first agent.

For a guided start

Let's talk about your case

Thirty minutes with us to scope a first use case and its gain.

To audit

Read the code

Apache 2.0. Your technical team can read it all, install it all, keep it all.