Concepts

These terms have the same meaning everywhere they appear in the product and in these docs.

Identity & access

TermDefinition
IdentityA unique entity in the platform — a user, an agent, a runner, or an app. Every identity has a type and a platform-wide ID.
UserA human identity provisioned on first OIDC login. Users belong to organizations and authenticate through the configured OIDC IdP.
OrganizationA multi-tenant boundary grouping users, agents, models, secrets, runners, and apps. Access control is scoped to organizations.
Cluster adminA platform-wide role with administrative access to all organizations, cluster-scoped runners, and platform users.
Organization ownerA per-organization role with full administrative access to the organization's resources.
Agent roleA grant from an identity to a specific agent: owner (manage roles, change availability, delete, edit config, start conversations), maintainer (edit config, start conversations), or participant (start conversations only). Org owners hold owner-level capabilities on every agent.
API tokenA long-lived credential for programmatic access. Tokens resolve to the same identity as their owning user. Format: agyn_<44 chars>.

Conversation

TermDefinition
ConversationA persistent exchange between participants (users, agents, or both). Conversations have a lifecycle (open → resolved) and accumulate messages over time. Stored as a thread internally.
ThreadThe internal storage record behind a conversation. Stores messages, participants, and acknowledgments.
ChatThe platform's communication interface. Users create conversations with any combination of users and agents in a single list-detail view.
ParticipantA user, agent, or app taking part in a conversation. Any participant can post messages; the participant's role determines who else can be added.
ReminderA scheduled follow-up attached to a conversation, created by an agent. Reminders notify the user at a specified time and can be cancelled.
Conversation statusUser-controlled lifecycle state: Open or Resolved.
Activity statusSystem-derived state for conversations with agent participants: Running (agent is processing), Pending (workload starting or retrying), or Finished (no unread messages or active workload).

Agent

TermDefinition
AgentAn AI entity configured with a model, runtime image, init image, tools, secrets, and instructions. Agents receive messages, reason, execute tools, and respond in conversations.
Agent availabilityWho may initiate conversations with the agent. internal allows any org member; private allows only identities holding an agent role.
ModelA platform-side identifier (e.g. gpt-4o) that maps to an LLM provider and a remote model name. Agents reference models by their platform name.
LLM providerA configured upstream model provider (OpenAI, Anthropic, etc.) with endpoint, auth method, and credentials.
SecretA credential value stored either locally (encrypted) or by reference in an external secret provider (e.g. Vault). Referenced from agents, MCP servers, hooks.
VolumeA persistent disk attached to an agent or sub-resource. Agents use volumes for working memory and durable state.
HookAn event-driven function attached to an agent. Runs as a sidecar container in response to platform events.
SkillA reusable prompt fragment placed on the agent's filesystem at startup.
Init scriptA shell script run before the agent CLI starts. Used for environment preparation.

Run & observability

TermDefinition
RunA single execution cycle within a conversation, triggered when the agent processes unacknowledged messages. A conversation accumulates multiple runs over its lifetime.
Run eventA discrete step within a run: a message received, an LLM call, a tool execution, or a context summarization. Events are the atomic unit of observability.
WorkloadA running pod attached to a conversation's agent. Lives only while the agent is processing or within the idle timeout.
ContextThe set of items (messages, tool results, memory, summaries) assembled into a prompt for an LLM call. Inspectable per LLM event.
SummarizationA run event where the agent's context is compressed to stay within token limits.
TraceThe recorded sequence of events for a run. Captures full LLM call context, tool inputs/outputs, and timing.

Tools

TermDefinition
MCP serverA Model Context Protocol server that exposes tools to agents. Runs as a sidecar in the agent pod, accessed over localhost.
ToolA capability available to an agent via an MCP server. Accepts structured input, executes an action, returns output. Produces stdout/stderr streams.
files-mcpA platform-provided MCP server that lets agents read uploaded files by ID.

Runtime & infrastructure

TermDefinition
RunnerA registered execution environment that hosts agent workloads. Cluster-scoped runners serve all organizations; org-scoped runners serve one organization.
WorkloadA running agent process — pod + sidecars + volumes — provisioned on a runner.
ContainerAn individual container within a workload, accessible via terminal logs.
DeviceA user device enrolled into the platform's private network, used to reach exposed services in agent containers.
Port exposureA reachable endpoint for a service running in an agent container. URL form: http://exposed-<id>.ziti:<port>.
NotificationA real-time event delivered to the UI via WebSocket. Drives live updates for conversations, messages, runs, and tool output.

Apps

TermDefinition
AppAn independently deployed service that interacts with conversations on behalf of external systems or platform capabilities.
App installationAn organization's activation of an app, with its own configuration and permissions.
Reminders appA platform-provided app that lets agents create scheduled follow-ups in conversations.
Telegram connectorAn app bridge between Telegram and Agyn conversations.