Tooling Layer
CLI tools, MCP adapters, skills, specs, AutoSkills, and patterns for creating new tools.
The tooling layer converts repeatable tasks into callable capabilities. Some are local CLI-first tools; others are external service adapters exposed through MCP. In both cases, the goal is to keep the agent from depending only on generative reasoning.
Local Agent Tools
The local tools toolkit acts as the base layer. It includes:
tools.jsonas the tool index;README.mdas the general guide;AGENTS.mdwith rules for agents;SKILL.mdfor toolkit activation;stack-template/bootstrap.pyfor creating new tools;- auxiliary usage telemetry tools;
- documentation for the background research worker.
MCP Adapters
Not every tool is a local CLI utility. Some capabilities are exposed through MCP adapters, especially when the system needs to call external research services. This keeps provider-specific code behind a stable tool boundary and lets agents treat external research as another callable capability.
Standard Tool Structure
For local CLI utilities, the standard structure is:
Each file has a role:
| File | Purpose |
|---|---|
main.py | Executable CLI entry point. |
lib.py | Domain logic when the tool grows. |
requirements.txt | Minimal Python dependencies. |
package.json | AutoSkills integration. |
pyproject.toml | Modern Python configuration. |
README.md | Human-facing usage. |
SKILL.md | Agent-facing usage: when to activate it and how to run it. |
TOOL.md | Technical specification written before implementation. |
tests/ | Basic or advanced verification. |
Creation Flow
Media Automation Tools
The media toolkit applies the same pattern to a concrete domain: movie and series search and management.
| Tool | Function |
|---|---|
torrent-finder | Search sources by title, year, quality, or codec. |
subtitle-finder | Search subtitles. |
dubbed-finder | Detect releases with dubbed audio. |
media-info | Extract technical video metadata. |
media-organizer | Organize downloaded media. |
Why CLI-First
A CLI-first approach provides:
- reproducibility;
- testability;
- execution by humans and agents;
- observable logs and errors;
- simple integration with other workflows;
- less dependence on visual interfaces.
Main Lesson
An agent tool should not be only code. It should include its discovery path, usage contract, agent context, and verification strategy.