Connect the right people.
Organization roles and host visibility define who can reach each machine and its work.
Organization roles / host visibilitySecurity at Glueprint
Give your agents room to work, with a security model your team can inspect. Execution stays on your hosts. Access is governed by explicit checks.
Execution, content transport, and key delivery are separate concerns. Explore the key path your team uses.
Repositories, agents, terminals, and files.
Plaintext while in useCiphertext, routing metadata, and presence.
Wrapped key storageKey delivery via AWS KMSInside the key-delivery boundaryDesktop, browser, and mobile.
Plaintext after authorizationThe relay stores wrapped key material and does not receive the credential-derived key that unlocks it. A database or relay compromise alone does not provide the client-unwrapped key.
For an authenticated SSO session, the relay asks AWS KMS to unwrap the tenant-bound key and delivers its bytes to the client over TLS. An active relay compromise is inside this trust boundary; a database snapshot alone remains insufficient.
Encryption protects content in transit. Server-side authorization decides who can receive it, who can act, and which actions need a person.
Organization roles and host visibility define who can reach each machine and its work.
Organization roles / host visibilityTeam membership, board roles, and object ownership govern collaboration. Task Flow Definitions restrict transitions by role.
Viewer / contributor / leadPermission tiers, tool filters, quiet hours, and supervision rules shape what assistants can do.
Permissions / tools / supervisionEnterprise supports SAML, OIDC, and SCIM 2.0 for identity-provider login and lifecycle provisioning.
Identity controlsSecurity-sensitive account, team, task, and administrative activity is recorded in audit events.Explore access controls
A useful security model makes its limits visible. Here’s what sits alongside content encryption.
Read about endpoint storageTenant, host, session, assistant and team identifiers; presence; sequence numbers; timestamps; event labels; and selected fixed-enum/read controls remain visible for routing.
Authorized hosts and clients decrypt content to execute or display it. A compromised endpoint can expose plaintext and key material available there.
The coding agent and model provider you configure receive prompts and context directly, under their own security and privacy terms.
The tenant Data Encryption Key is shared within the organization. Host visibility, membership, roles, and object permissions control delivery and access.
Encryption primitives, credential storage, authorization, and infrastructure. Open a section to inspect the implementation model.
Glueprint uses a tenant Data Encryption Key (DEK) for relay content. Password, recovery, and agent-host paths store and transport only wrapped key material. Enterprise SSO stores a second KMS-wrapped key and permits the relay to call AWS KMS Decrypt for an authenticated SSO session.
Authenticated encryption protects confidentiality and detects ciphertext modification.
Every encrypted message generates a fresh nonce for GCM nonce uniqueness.
A password-derived KEK unwraps the tenant DEK on the client.
An API-key-derived KEK unwraps the tenant DEK on the host.
AWS KMS unwraps a tenant-bound key; the relay returns its bytes to the authenticated client over TLS.
Raw DEK bytes may be cached per tab under a user-and-tenant scope so a reload can recover the active key.
The desktop keeps its active relay DEK in zeroizing process memory and clears that copy when the encryption manager locks. Before host registration, browser sign-in supplies a separate temporary DEK held only in desktop process memory; it is cleared after activation, on logout or a new sign-in, or on the first use after its 30-minute lifetime expires, and it does not survive a restart. The portal may keep exportable key bytes in tab-scoped session storage until logout or the browser session ends. Device compromise can expose plaintext or key material available on that endpoint.
Human and agent connections use separate credential types and scopes. Desktop sign-in uses PKCE, human relay clients use JWTs, and agent hosts authenticate with API keys.
User tokens and API keys are kept in separate files under the Glueprint data root; Unix writes use owner-only permissions.
Email verification, optional MFA, rotation, and refresh-family revocation reduce account-token risk.
Activation binds the connection to an organization and host identity; an environment-variable override is also supported.
Claimed domains, identity-provider login, and lifecycle provisioning support managed access.
Server-side authorization evaluates organization role, host visibility, team membership, board role, and object ownership before data or control requests are accepted. Sensitive WebSocket controls pass through explicit subtype allowlists and role gates.
Repositories, agent processes, terminals, and the primary operational store remain on the selected host. Desktop and daemon data default to ~/.glueprint/, with YAML and JSONL files for configuration and event history.
Rust services expose typed command boundaries to the desktop frontend, and shared storage paths use atomic-write patterns where implemented. Tokens, API keys, and managed harness keys use dedicated local files rather than universal OS-keychain storage; Unix writes use owner-only permissions. Endpoint account controls, filesystem permissions, and disk protection remain part of the security boundary.
Authenticated HTTP and WebSocket connections carry control and encrypted event traffic. Sequence tracking and gap fill support ordered delivery and recovery; they are delivery controls rather than a replacement for cryptographic authentication.
Glueprint does not claim a certification, penetration-test result, or security property that has not been independently established.
Architecture reference last updated .
Reviewing Glueprint for your organization, or found something we should investigate? Talk directly with us.
security@glueprint.aiWe acknowledge and assess reports as promptly as practical and coordinate remediation. If an incident affects personal data, we investigate, contain it, and notify affected people and authorities as required by applicable law.