Skip to content

Security and credential boundaries

Nori’s security model separates the organization broker, the session machine, the agent process, and external integrations. Access is granted through explicit organization configuration rather than inferred from a prompt.

Each hosted organization has a single-tenant broker at its own <org>.norisessions.com subdomain. The broker owns organization membership, fleet configuration, integration records, session metadata, and credential distribution for that organization.

Session machines are discovered and managed within the active provider’s organization or deployment boundary. Provider-specific labels, identities, and network rules prevent the broker from treating unrelated machines as members of the fleet.

One claimed machine serves one active session. The environment is supplied by the configured compute backend:

  • Modal uses a managed gVisor sandbox by default, with a separately configured VM Sandbox option for eligible organizations.
  • Sprites provides the legacy hosted VM backend.
  • AWS BYOC uses private EC2 instances and broker-restricted network access.
  • Local Docker uses labeled containers on a private deployment network.

Capabilities differ by backend, so this documentation does not reduce every environment to one generic “secure sandbox” claim. See fleets and compute for the support boundary.

Credentials are configured in the broker dashboard or through an authorized Nori management flow. Distribution depends on the credential type:

  • OAuth refresh credentials remain at the broker where the integration supports broker-side refresh. The broker sends refreshed access material to machines.
  • API keys and custom environment values are written into protected session environment files or agent-specific configuration files.
  • GitHub App access uses installation-scoped credentials. A configured personal access token takes precedence when both GitHub methods are present.
  • Google Workspace access is per user and follows the scopes granted during its OAuth connection.
  • Managed MCP connections use the access mode and authorization supported by the selected server.

Credential changes are distributed to running machines. Some OAuth integrations also refresh on a cadence and re-push new access tokens.

Administrators can define arbitrary environment variables for sessions and mark values as admin-only. These values are useful for tools that do not have a dedicated Nori integration, but they may be long-lived secrets. Teams should scope them narrowly and rotate them according to the external service’s policy.

Slack has two separate control planes:

  1. Who may drive the bot: open, guarded, or strict workspace access.
  2. How far the agent may reach into Slack: channel-only, public-member, or direct-token access.

The default agent-reach mode is channel-only. It proxies Slack operations and limits them to the session thread’s channel. Direct-token mode places a raw Slack token in the session and is intended only for internal environments.

Optional scope groups for direct messages, file upload, and file download can be disabled before the Slack app is installed.

The selected agent has its own permission and tool behavior. Nori supplies the machine and connected credentials, but an agent can only use what reaches its environment. The Nori CLI separately supports approval and sandbox policy for local use.

Teams should treat repository permissions, MCP access modes, infrastructure credentials, and custom environment variables as independent grants. Connecting one integration does not authorize another.

Session machines are not permanent storage. Commit repository changes and store important non-repository outputs in an appropriate connected system. Transcripts and checkpoints support continuity, but they do not turn the working directory into an unlimited archival volume.

  • Use the narrowest repository and integration scopes that support the task.
  • Prefer GitHub App installation access over broad personal tokens.
  • Keep Slack agent reach at channel-only unless a broader mode is necessary.
  • Review custom environment variables as secrets, even when their values are hidden in the dashboard.
  • Explicitly release completed sessions.
  • Verify the compute and agent support matrix before designing a regulated or customer-specific deployment.

Continue with integrations and credentials or Slack access control.

this page as markdown: /sessions/concepts/security-and-credentials.md· index for agents: /llms.txt