Skip to main content

Sovereignty

Autoheal is built for regulated environments, where you control where agents run, what they can access and what they can change. This page covers the deployment models, including connected and airgapped BYOC, and the governance controls that apply to all of them.

For platform security posture, including encryption, certifications, vulnerability management, and incident response, see Security.

Deployment models​

Autoheal offers two deployment models. Governance controls, authorization boundaries, and approval policies are identical in both. The difference is where the platform runs and where your data comes to rest. See Architecture for how the platform is built.

Fully-managed SaaSBYOC (Bring Your Own Cloud)
AvailabilityDefaultPaid add-on
Runs inAutoheal's cloudYour own cloud account (connected) or hosts you provision (airgapped)
Data at restAutoheal's infrastructure, logically isolated per tenantYour infrastructure, inside your network
LLM inferenceAutoheal-managed provider accounts, with models selected per taskYour model endpoint or provider account
UpgradesAutomaticYou control the upgrade window
Best forMost teams, and the fastest path to a first agent runStrict data residency, regulatory, or contractual requirements

Fully-managed SaaS​

The default. Autoheal runs the platform and you connect your tools. Your environment is logically isolated from every other tenant, your data never crosses a tenant boundary, and you get upgrades as they ship. This is the fastest path to a first agent run and the right choice for most teams.

BYOC (Bring Your Own Cloud)​

Available as a paid add-on. See BYOC for the deployment architecture. The entire Autoheal stack is deployed inside your own cloud account. As a result, all data, including investigations, memories, skills, and agent run traces, stays contained in your environment and never leaves it. You control the upgrade window. LLM inference can be routed through your own provider accounts, so prompts and completions stay within your contractual boundary too.

BYOC runs in one of two modes:

ConnectedAirgapped
Runs onYour Kubernetes cluster (Amazon EKS or Azure AKS)Docker Compose on hosts you provision, on an isolated network
Software deliveryRelease images pulled into your account for each upgradeA versioned bundle of every image, with checksums, that you import. Nothing is pulled from a registry at runtime
Operator accessOptional: a scoped, revocable role for the install or upgrade windowNone. Your team installs and upgrades from the bundle
LLM inferenceAn in-account model or a direct provider you chooseA model endpoint reachable from your isolated network, which you configure

Choose connected BYOC when data residency or a contract keeps data in your account but the environment can pull releases. Choose airgapped when the environment has no path to the internet or to Autoheal at all. In that case, accept that upgrades become a bundle import your team schedules and runs.

BYOC is the right choice when a data residency requirement, a regulatory regime, or a customer contract means production telemetry cannot leave your infrastructure. Contact support@autoheal.ai to enable it.

Agent governance​

Governance controls apply identically in both deployment models.

Authorization boundaries​

Every agent has an explicit allow-list of integrations it may reach. Nothing outside that list is reachable. The platform enforces the boundary, not instructions in a prompt. As a result, an agent cannot talk itself into a data source it was not granted.

Each agent's tool filters decide which tools it may call: system_tools for Autoheal's own tools and integration_tools for the tools of its integrations. Both can treat reads and writes separately. An agent with no filters can call every tool its run can see, writes included. For least privilege, write operation_type == "read" and add a write only where the agent posts a result back, such as a comment on a pull request. Many integrations ship only read tools, so an agent has no way to write to them at all. Tool filters do not limit the agent's shell, which uses the project's CLI connections. An allowed integration write runs without pausing unless a governance policy requires approval for it. For this reason, most organizations start with a policy that gates every write operation. Policies match on the integration a call touches, so they do not apply to system tools. See Build Your Own Agents for how tool filters are written.

Approval policies​

Approval policies are set by admins and apply tenant-wide. Each policy matches on the integration's type, name, labels and operation type. A tool call that matches any policy pauses until a person approves it. A policy on operation type is write gates every write, including calls the platform can't classify, which are treated as writes. With no policies configured, tool calls run without approval. A run that needs a decision pauses as waiting for input instead of acting unattended.

Autoheal posts status updates, such as a ServiceNow work note or a Jira comment, to the ticket that started an investigation. The platform sends these updates, not an agent tool call. For this reason, approval policies don't apply to them. See Governance Policies for how approval policies are written and evaluated.

Audit​

Every agent run is recorded with its trigger, the integrations it queried, the context it loaded, each step it took, and the conclusion it reached. Combined with the evidence behind each conclusion, this gives regulated environments a complete and queryable record of what an agent did and why. Every tool call and approval decision is queryable in the Audit Trail. See Roles & Permissions for access control over these records.

Private evaluations​

The Evaluator scores agents on your own runs and outcomes, in your isolated tenant (SaaS) or your own environment (BYOC). No other customer's data shapes how your agents are scored or which context changes the Healer proposes.

Cost governance​

LLM inference spend is governed alongside access. Autoheal tracks model spend per agent and per team. Platform owners can then see which agents are expensive. They can set the dial between cost and autonomy rather than discover the tradeoff on an invoice. Frontier models handle complex reasoning, and lower-cost or open-weight models take the steps where evaluations show they perform well.

Under BYOC, inference routed through your own provider accounts appears on your existing cloud bill and within your existing commitments. See Healer for how per-run cost is reduced, including routing tasks to cheaper models.

Data residency​

Under fully-managed SaaS, data residency options are available for storage location. See Security. Under BYOC, residency follows your own cloud account, so data comes to rest in the regions you deploy into and nowhere else.

In both models, your data is never used to train AI models without your explicit consent.