How Helmr keeps agent work in bounds.

Agents work with real repositories, credentials and internal systems. This page explains what Helmr isolates, who can act, what it records, and where its guarantees stop.

01 · Isolation

Where agent work runs

Agents and their commands run inside a Linux microVM on a worker. The computer is defined in code (image, tools, resources, secret bindings and a prepare step), so what runs is reviewable and the same for every session.

Builds treat package install scripts as untrusted input. They run isolated, receive only the build secrets you select, and get no host Docker socket or unrestricted host filesystem.

Between environments
Separate authority, secret bindings and audit scope.
Inside one environment
Agent code is mutually trusted. Sessions that share a computer share its files and processes.
Security concepts →

02 · Credentials

Credentials

Bind a secret as protected and the VM holds only a placeholder. Helmr substitutes the real value at the network boundary for HTTPS origins you allow, and fails the request once the secret is revoked. Reading a secret returns its metadata, never the value.

Raw environment variables and files expose values to processes in the computer; use them only where a tool cannot work through a protected binding.

Secrets →

03 · Access

Who can act

Roles are sets of permissions granted to members, groups and API keys at organization, environment or single-agent scope. Standard roles cover common jobs; custom roles hold exactly the permissions you choose.

Members and groups
Each member’s access shows every permission and the assignment it comes from.
API keys
Belong to one environment and get nothing until a role is assigned.
Organization owner
Ownership alone grants no ordinary access, including audit access.
Answers to agents
Answering an agent’s question requires authority, and Helmr records who answered.
Permissions →

04 · Audit

What is recorded

Helmr records the operations it enforces, with the identity that acted, how it signed in and every authority check it evaluated. Refused requests are recorded too.

Recorded

  • Sign-ins and administrative changes
  • Session, turn and command transitions
  • Secrets made available to a computer
  • Refused requests and the reason
  • Reading execution content or audit events

Never recorded

  • Prompts and messages
  • Command arguments and output
  • File contents
  • Tokens and secret values

Who can read it

Reading audit events needs its own permission, granted with the Auditor role for the organization or an environment. Every search is itself recorded.

How long it is kept

Retention is a finite period the operator sets; there is no unlimited mode. The serving process can only add and read audit records; removing expired ones is a separate maintenance command.

Audit events →

05 · Deployment

Where Helmr runs

Run the control plane and workers in your own AWS account. The private reference keeps the Console, CLI and API on an internal load balancer, sends platform traffic through a proxy you maintain, and connects to Slack outbound, with no public endpoint.

Private deployment →

06 · Limits

What Helmr does not cover

Knowing where a guarantee stops matters as much as the guarantee.

Agent tool internals
Activity inside a native agent harness that does not pass through Helmr is not in the audit log.
Untrusted code side by side
Agents in one environment trust each other; Helmr does not isolate them from one another.
Log export
The audit API is for investigation, not a loss-free collector or archive.
External side effects
Retries reconcile delivery; they do not make external actions happen exactly once.
Erasure
Expired records are hidden immediately; secure physical erasure is not promised.

07 · Open source

Read the source

Helmr is Apache 2.0. Every boundary on this page is in the source, so you can verify it before you trust it with your repositories.

helmrdotdev/helmr on GitHub →