Skip to main content
There is no skills registry in the Mirobody C++ engine. The SKILL.md + metadata.json “Agent Skills” mechanism — a skills/ directory auto-discovered at startup — was part of an earlier design and is not implemented in the current C++ engine (Mirobody). The single agent in res/agents/base.cpp builds a fixed inline system prompt, picks an LLM client by provider, and streams the turn back, calling built-in MCP tools. It does not load SKILL.md files, and there is no skill loader.

​
What to use instead

The way to extend what the agent can do is to add an MCP tool — a small C++ file in res/mcp_tools/ that self-registers at compile time. Tools are “code the agent calls”: the model invokes one like a function and reads the return value. There is no separate “prose the agent loads into context” layer.

Adding Custom Tools

Add a C++ tool in res/mcp_tools/ with the MIROBODY_REGISTER_TOOL macro, then rebuild.

Built-in Tools

The 9 MCP tools that ship with the engine.

​
Guiding the agent’s behavior

If you want to change how the agent behaves — its tone, the domains it knows about, or when it reaches for a tool — the levers in the C++ engine are:
  • The system prompt. The agent’s instructions are a fixed inline string built in res/agents/base.cpp (build_system_prompt). It already tells the model to fetch uploads via list_files / read_file and health records via family_health. Editing that string and rebuilding changes the agent’s standing instructions.
  • Tool descriptions. Each tool’s description and per-Param text (declared in its res/mcp_tools/*.cpp) is what the model sees when deciding whether and how to call it — so a clear description is the main way to steer usage. See Built-in Tools for examples.
Changing the prompt or adding a tool is a compile-time change: edit the .cpp, then rebuild with ./build.sh. There is no runtime drop-in for either.

​
Next steps

Tools & Agent Overview

How the single agent, MCP tools, and providers fit together.

MCP Integration

Expose the tools to Claude, Cursor, and ChatGPT over /mcp.