Mellow vocabulary

A plain-language map of the concepts used throughout the app and this reference.

Use these definitions to distinguish the choices in a conversation and the boundaries behind them. Terms are grouped by the work they describe rather than by where they happen to appear in a settings menu.

Work and organization

Agent

A saved role with instructions, model preferences, capabilities, and permission choices. An agent can participate in several projects; a project can contain chats from several agents.

Orchestrator

The built-in Mellow coordinator. It inspects and changes supported configuration through review, then delegates hands-on work to allowed specialists.

Project

A topic-based group of conversations with common instructions, knowledge collections, memory scope, and an optional default folder and agent.

Workspace

A shared service context with account and membership boundaries. In Mellow, distinguish a Cloud workspace from a local project and from a paired device.

Working Folder

The host directory selected for a conversation or supplied as a default. It identifies the file context for supported operations and is separate from Sandbox execution.

Prompt

The message or instruction given to a model. A useful task prompt names its inputs, expected result, boundaries, and evidence of completion.

System prompt

Persistent context and instructions assembled for a model request. Agent configuration, project guidance, and runtime information can contribute to it.

Artifact

A file explicitly surfaced as a deliverable in a conversation. A file can exist on disk without having been shared as an artifact card.

Activity log

A record of events such as task progress or tool execution. Inspect the event and result rather than treating the presence of an entry as proof of success.

Insights

The app area for inspecting runtime activity and diagnostic information. It helps explain which capabilities were used and where a run failed.

Capabilities and context

Tool

An executable operation with an input schema and result. The runtime controls availability and permissions; a model mentioning an operation does not mean it has performed it.

Skill

An authored playbook with instructions and optional reference files. Skills guide work but do not independently grant tools or access.

Method

A stored reusable procedure with dependency and outcome metadata. Loading a method supplies guidance; it does not certify that the procedure will succeed.

Plugin

A package that contributes supported capabilities or content. Installation, external authentication, and permission to execute are distinct setup stages.

MCP

Model Context Protocol: a connection format through which a server exposes tools or related resources. The discovered count reflects that server and account scope, not every capability in Mellow.

Knowledge

Document collections you explicitly make available to an agent or project. Retrieval selects relevant material from those sources.

Memory

Retained conversational context used for continuity. It differs from source documents in Knowledge and deliberately structured records in an agent database.

Embedding

A numerical representation used to compare semantic similarity. It assists retrieval but does not establish that a returned passage answers the question correctly.

Core Model

The model selected for background processing such as memory extraction. Its readiness and locality can differ from the model in the active chat.

Subagent

A specialist running a bounded task for another agent. It uses its target configuration and returns a result to the parent conversation.

Delegation

Assigning a task to an allowed specialist. Clear inputs, ownership, and expected results make delegated work easier to combine and verify.

Approval

A review point for a proposed operation. Allowing a request permits the operation to proceed; its result must still be checked.

Cron

A compact timing expression used for schedules. Inspect the next-run preview to confirm how an expression is interpreted.

Models and requests

Model

The system that generates a response or a proposed next step. Different models have different context limits, modality support, resource needs, and service requirements.

Local model

A model whose inference runs on the Mac. External tools used during the conversation can still make network requests.

Cloud model

A model served remotely by a provider. The provider receives the content included in its request, which can include retrieved context and tool results.

Provider

The configured service or runtime that supplies models. Authentication to a provider is separate from membership in a Cloud workspace.

Context window

The amount of input and generated context a model can handle for a request. Long chats may need selective retrieval or compaction.

Token

A unit used to measure model input and output. Tokens are not the same as words, and conversion varies across text and models.

Temperature

A generation parameter affecting sampling variability where the selected model supports it. A lower value is not a guarantee of factual correctness.

Parameters

The learned numerical values in a model. Parameter count is one size indicator, not a complete measure of quality or memory requirements.

Quantization

A representation that reduces numerical precision to lower model size or resource use. Compatibility and quality depend on the specific model format.

Apple Intelligence

Apple-provided system AI capabilities. Availability depends on the Mac, operating system, and system configuration.

Apple Silicon

Apple-designed Mac processors. Check the application and model requirements for the supported architecture.

MLX

A machine-learning framework used by supported local runtimes on Apple silicon. Runtime support does not imply that every downloaded model format is compatible.

Hugging Face

A model and dataset hosting service used by many publishers. Inspect a model repository’s format, requirements, and license before choosing it.

Transcription

Converting speech into text. Successful recognition, inserting that text into a field, and synthesizing speech are separate stages.

Wake word

A configured spoken trigger for a voice interaction. Its recognition and microphone requirements are separate from ordinary typed chat.

Connections and access

Endpoint

The address of a service operation. Reachability, authentication, and permission to perform a requested action are separate checks.

API

An interface for programmatic requests. Consult the operation schema and response rather than assuming a chat feature is available through every endpoint.

API key

A credential used to authorize requests to a particular service. Treat it as private and configure it only for the intended provider or endpoint.

Access key

A Mellow credential granting a defined access scope. Current Mellow-format keys use the sk-mellow prefix. A host URL alone does not provide the same authorization.

Public Link

A reachable address for a shared host or service. Public reachability does not imply anonymous access, a browser chat interface, or unrestricted control.

Loopback

A network address that refers to the current computer, such as 127.0.0.1. On a phone, that address points to the phone rather than your Mac.

Relay

A transport intermediary used to connect endpoints. Relay availability does not prove the destination agent is online or authorized for your request.

Secure Channel

An authenticated protected connection between participants. Verify the intended host identity during pairing instead of relying on a display name alone.

End-to-end encryption

Protection in which intended endpoints hold the keys needed to read content. It does not prevent an authorized endpoint or selected model service from processing data deliberately supplied to it.

Identity

The app or agent identity used in supported access and sharing flows. Local identity readiness and the availability of hosted identity services are different states.

Recovery phrase

Secret recovery material for an identity where that mechanism is supported. Keep it private and use only the product’s designated recovery flow.

Mellow Cloud

The user-facing cloud connection area. Connected workspaces, published agents, and model-provider services can have separate availability and permissions.

Mellow Router

A hosted routing component where configured. Its absence in a build does not mean the local model runtime is unavailable.

Credits

A service-specific usage balance where a hosted product uses one. A loading or authentication error should not automatically be interpreted as insufficient credits.

Privacy and local execution

Sandbox

The supported isolated execution environment for agent programs. It is separate from a host Working Folder; its implementation and capabilities depend on the operating system.

Privacy Filter

A configured processing stage for recognized sensitive content in supported outbound requests. Inspect its scope and detector behavior; it is not universal proof that every secret was removed.

PII

Personally identifiable information, such as contact details or identifiers that can identify a person. Detection quality depends on format, context, and the enabled detector.

Keychain

macOS secure credential storage. Do not erase entries as a routine troubleshooting step when you intend to preserve sign-ins and device identity.

FileVault

macOS disk encryption. It protects data at rest under its operating assumptions and is separate from application-level database encryption.

Full Disk Access

A macOS privacy permission for access to protected data. Grant it only when the intended feature requires it; it is broader than selecting one working folder.

Automation permission

macOS consent allowing one application to control another through supported automation. It is separate from microphone, Accessibility, and folder permissions.

JSON YAML

Structured text formats used for configuration and data. Formatting and field names matter; consult the current schema before applying a document.