Skip to main content

Core concepts

Ujex is a trust layer under your agent loop. These are the nouns used across the dashboard, CLI, SDKs, and audit records.

Agent​

An agent is a first-class identity owned by a human account. Mailboxes, device keys, secrets, tool connections, memory files, usage events, approvals, and evidence all hang off the agent.

Owner​

An owner is the human liable for the agent. Owners create agents, approve risky actions, rotate/revoke keys, register secrets and tool connections, and export evidence.

Device key​

A device key is the credential an agent process holds. It is shown in plaintext once, stored only as a hash, and exchanged for a Firebase custom token with an agent_id claim. Rotate or revoke it from Settings.

Scope​

A scope is a named permission on a device key, such as postbox, audit, memory, tools, or secrets. Callables check scopes before doing work.

Agent Passport​

The Agent Passport is signed JSON for an agent identity. It includes owner hash, mailbox metadata, active scopes, active key prefixes, delegation summary, revoked counts, and latest audit anchor. It does not expose raw owner email.

Approval​

An approval is a human decision for a risky action. Approval-gated sends store an action digest so the decision and the action can be matched later.

Evidence​

Evidence is the exportable proof set: hash-chained audit records, manifest.json, optional manifest.sig, latest anchor data, and trust-bundle.txt. The CLI verifier checks the chain and manifest signature.

Secret​

A secret is encrypted server-side with KMS. Owners can store, list metadata, and delete secrets, but they cannot reveal plaintext in the browser. Agents with the secrets scope can read the value at runtime by passing the current sessionId; Ujex marks that session as sensitive_access=true.

Tool connection​

A tool connection is provider-agnostic SaaS access metadata plus encrypted token references. Agents invoke tools without seeing raw credentials.

Guardrails​

Guardrails are policy, risk, prompt-injection, exfiltration, approval, and estimated-spend decisions before an action runs. Spend is estimated action cost, not reconciled provider billing.

Memory provenance​

Memory provenance is markdown-first long-term context with source labels. Legacy recall* APIs remain for compatibility; new integrations should use the memory* surface.