Repository navigation
Glossary
The words this wiki uses, in alphabetical order.
activate_skill: The tool the model uses to load a skill's instructions, or one of its files. Dive offers it whenever at least one skill is switched on. See Tools and skills.
Agent Mode: A per-mode setting that raises the tool budget to 25 calls by default and tells the model to plan before it works. See How the model uses tools.
Agent Skill: The full name of a skill, from the Agent Skills format. See skill.
Allowed folders: The folders the code tools can read: ~/dive/workspace, plus any listed in ~/dive/allowed-dirs.json. See Native tools.
Custom tool: A tool you define in Settings as a shell command or a piece of JavaScript. It belongs to one mode. See Create a custom tool.
Data folder: ~/dive, where Dive keeps everything: conversations, settings, notes, skills, external tools and the library. See Data folder and settings.
DATABASE switch: The switch in the side panel that grounds answers in passages from your indexed library, also called Database Context. While it's on, the model gets no tools in most modes.
Dive modes: llama.cpp, LM Studio, Ollama and Cloud: the four modes where Dive supplies the tools and skills. Pi mode is the exception.
Execution Trace: The collapsible list of every tool call in a reply, marked when something failed.
External tool: A tool that you install as a plugin: a Node.js module in ~/dive/plugins. See Create an external tool.
Forced tool: A tool run through a slash command, such as /calc, instead of chosen by the model. See Slash commands.
INPUT button: A button above the message box that types a tool's slash command, turned on in the INPUT column of Settings > TOOLS.
Lesson: A standing instruction added to one mode's system prompt, written by you or saved by the model with remember_lesson. Lessons live in Settings > PROMPTS > LESSONS.
MCP: The Model Context Protocol, a standard way for programs to offer tools to AI apps.
MCP server: A program that offers tools over MCP. Dive starts the ones you list in Settings > TOOLS > MCP SERVERS. See Connect an MCP server.
Mode: What Dive talks to: llama.cpp, Pi, Cloud, LM Studio or Ollama. See Choosing a mode.
Native tool: One of the 29 tools that come with Dive. See Native tools.
Native tool calling: Sending tools to the model as structured function definitions, which it answers with structured tool calls. The alternative is the text prompt.
Permission prompt: The dialog that asks before a risky tool runs, with ALLOW, DENY and CANCEL. See How the model uses tools.
pi: A coding agent that runs in the terminal. Pi mode puts Dive in front of it. See Pi mode.
pi extension: Code that adds tools or commands to pi. Dive's own tools don't reach pi, so this is how Pi mode gets new tools.
Plugin: The code form of an external tool. The two words mean the same thing in this wiki.
Router: A llama-server started in --models-preset mode, which loads models on request. Dive can use one instead of starting its own server. See llama.cpp mode.
Runtime folder: The copy of Dive's server code that the desktop app runs, refreshed on every launch. On macOS it is ~/Library/Application Support/dive-desktop/runtime.
Skill: Written instructions the model follows for one kind of task: a folder with a SKILL.md file. A skill runs nothing by itself. See Create a skill.
SKILL.md: The file at the heart of a skill: YAML frontmatter with a name and a description, followed by the instructions.
Slash command: A command at the start of a message, such as /wiki or /skill:name, that tells Dive exactly what to do. See Slash commands.
Text prompt: Describing the tools in the system prompt, so the model calls one by writing <call:tool_name>{...}</call>. Cloud mode always works this way, and the local modes fall back to it.
Tool: An action the model can take. Dive runs the tool's code and gives the result back to the model. See Tools and skills.
Tool budget: How many rounds of tool calls one request may use: 6 in llama.cpp, LM Studio and Cloud mode, 10 in Ollama mode, and 25 by default with Agent Mode.
Workspace: ~/dive/workspace, the folder where file_operations reads and writes and where fetch_paper saves papers.
These pages describe Dive 5.0.8. Source code: DuckTapeKiller/dive. Spotted a mistake? Open an issue.
Get started
Tools and skills
Build your own
Share and stay safe
How Dive works
Help