Glossary
Definitions for the resources and concepts in Orca Agent Engine, with the canonical spelling of each term.
The terms these docs use, and how each one is written. Where a term names a registry resource, the entry links to the page that owns it.
Most terms are ordinary lowercase nouns. Workspace is capitalized because it names a specific
StreamNative Cloud resource rather than a generic one, and legacy feature names such as
Agent Function keep their capitals.
- agent
- A versioned registry resource declaring the model, system prompt, tools, MCP servers, and skills that shape how a session behaves. Not "Agent".
- Agent Function
- The deprecated function-mesh code unit that ran an agent. Superseded by triggers. Capitalized: it is a product feature name.
- AgentContext
- The API an Agent Function used to reach tools, state, and message metadata at runtime. Deprecated with Agent Functions.
- custom tool
- A tool your own application executes: the session pauses, hands you the call, and waits for a result.
- environment
- A reusable container template setting the packages and network policy a session executes under. Not "Environment".
- event
- A message exchanged with a session - a user turn, a tool call, a tool result, or a status change. Not "Event".
- file
- An uploaded binary asset a session can mount as a read-only resource.
- guardrail
- A policy rule that can allow, ask for approval, or deny an agent action at session, agent, Workspace, or organization scope.
- MCP server
- An external Model Context Protocol endpoint whose tools an agent can call. Declared inline on the agent, never a standalone registry resource.
- memory store
- Workspace-scoped storage that agents read and write across sessions, mounted into the sandbox under
/mnt/. - model price
- The effective token rates for a provider and model, used by the registry to account for session spend.
- permission policy
- Whether a tool runs immediately or pauses the session for approval -
always_alloworalways_ask. - provider
- An LLM backend a Workspace exposes, which an agent reaches through its
modelfield. Not "Provider". - registry
- The per-Workspace catalog of resources, and the API every CLI and SDK call goes through. Not "Registry".
- runtime
- The compute layer that executes sessions inside a Workspace. Not "Runtime".
- session
- A running instance of a pinned agent version inside an environment, driven by events. Not "Session".
- skill
- A versioned filesystem bundle containing a root
SKILL.md, which an agent loads on demand for domain-specific context. Not "Skill". - thread
- A server-created subdivision of a session. Threads are not created directly. Not "Thread".
- tool
- Something an agent can call beyond producing text, declared on the agent as a built-in toolset, an MCP toolset, or a custom tool.
- toolset
- A group of tools enabled together -
agent_toolsetfor the sandbox tools, ormcp_toolsetfor one MCP server. Not "Toolset". - trigger
- Automation that starts sessions on a cron schedule; StreamNative Cloud also accepts Pulsar and Kafka messages. Not "Trigger".
- vault
- Workspace-scoped credential storage. A session passes vault IDs, and the gateway injects the matching credential into MCP calls. Not "Vault".
- vault credential
- One secret inside a vault, bound to an MCP server URL. Kinds are
static_bearer,mcp_oauth, andenvironment_variable. - Workspace
- The StreamNative Cloud resource that hosts Orca and is the isolation unit for tenancy, identity, and quotas. Capitalized: it names a specific platform resource. Not "workspace".