Glossary
Key terms and definitions used throughout DotAgents.
| Term | Definition |
|---|---|
| ACP | Agent Client Protocol — protocol for agent-to-agent communication and delegation. Based on JSON-RPC 2.0. |
| Agent | An AI persona with specific skills, tools, and behavior. Runs internally or as an external process. |
| Agent Bundle | A portable package containing an agent's profile, skills, and configuration for sharing. |
| Agent Profile | The complete definition of an agent: identity, system prompt, tool access, skills, and connection. |
| .agents/ | The open standard directory for agent configuration. Works across DotAgents, Claude Code, Cursor, and others. |
| Context Budgeting | Automatic management of the LLM's context window, shrinking messages when limits approach. |
| Delegation | When one agent assigns a subtask to another agent via ACP. |
| Dictation Mode | Voice mode that transcribes speech and inserts text into the active application. |
| Elicitation | MCP 2025 protocol feature where a server requests additional input from the user during tool execution. |
| Emergency Stop | Ctrl+Shift+Escape — immediately aborts all active agent sessions and tool executions. |
| Frontmatter | YAML-like metadata at the top of markdown files, delimited by ---. Used in skills, memories, and agent profiles. |
| Global Layer | The ~/.agents/ directory — personal agent config shared across all projects. |
| Guidelines | Additional behavioral rules in an agent profile, supplementing the system prompt. |
| Hands-Free Mode | Voice Activity Detection (VAD) mode on mobile — listens and sends automatically without button presses. |
| Kill Switch | Emergency stop mechanism. Same as Emergency Stop. |
| Langfuse | Open-source LLM observability platform. Optional integration for tracing and debugging. |
| LLM | Large Language Model — the AI model that powers agent reasoning (e.g., GPT-4o, Llama 3). |
| Loop | A recurring task that runs on a schedule (e.g., "check email every 10 minutes"). |
| MCP | Model Context Protocol — Anthropic's open standard for connecting AI models to tools and resources. |
| MCP Client | The side that calls tools. DotAgents is an MCP client. |
| MCP Server | An external process or endpoint that exposes tools. Connects via stdio, WebSocket, or HTTP. |
| Memory | Persistent context stored in .agents/memories/ that persists across sessions. |
| Panel Mode | Compact floating window mode for the desktop app. |
| Parakeet | Local ONNX-based speech recognition model. No API key required. |
| Remote Server | Fastify HTTP server in the desktop app that mobile and external clients connect to. |
| Session | A single conversation thread with an agent, including messages and tool executions. |
| Skill | A markdown instruction file that teaches an agent how to do something. Stored in .agents/skills/. |
| STT | Speech-to-Text — converting spoken audio to text. Providers: OpenAI Whisper, Groq, Parakeet. |
| System Prompt | Core instructions that define an agent's behavior, personality, and constraints. |
| tipc | Typed IPC — type-safe communication between Electron's main and renderer processes. |
| Tool | A callable capability exposed by an MCP server (e.g., github:search_repositories). |
| Tool Approval | User confirmation required before a tool executes. Configurable per-agent and per-tool. |
| TTS | Text-to-Speech — converting text to spoken audio. Providers: OpenAI, Groq, Gemini, Kitten, Supertonic. |
| VAD | Voice Activity Detection — automatically detects when the user starts and stops speaking. |
| Workspace Layer | The ./.agents/ directory in a project — overrides global config for that project. |
Next Steps
- Architecture Overview — System design
- The .agents Protocol — Protocol details