Core AI Application Concepts Explained: Tools, Skills, MCP, Agents & Rules

Although AI’s power is undeniable, many treat it like a wish‑granting box and end up with unsatisfactory outcomes because they provide vague prompts; the article argues that precise descriptions are required to obtain useful results.

1. Tools & MCP

Local abilities are operations the computer can perform directly, such as executing Bash commands, file CRUD, web fetching, or network searches. In OpenCode these are called Tools , and any OS‑exposed capability can be invoked.

When local software or remote services conform to the Model Context Protocol (MCP) —an open, standardised communication protocol released by OpenAI in Dec 2024 for safe, bidirectional interaction between large language models and external tools—OpenCode can automatically control them.

MCP works like a USB‑Type‑C cable: any compliant tool can be connected. OpenCode supports two MCP server modes:

Local: a MCP process started on the user’s machine via the command line, suitable for local tools and scripts.

Remote: an HTTPS‑based MCP service, suitable for cloud services and third‑party platforms.

2. Rules

Rules define the standards an AI must follow within a project. For example, a development team can create an AGENTS.md file at the project root to declare package structure, code style, and workflow conventions. Global rules placed in ~/.config/opencode/AGENTS.md apply to all OpenCode sessions, and OpenCode searches from the local directory up to the global directory for these files.

~/.config/opencode/AGENTS.md

3. Skills

Skills are analogous to cooking recipes: they describe *how* to use the tools for a specific task. A Skill is typically a Markdown file (e.g., code_review.md) that does not list available tools but explains the step‑by‑step procedure—what to do, when to use which tool, and which parameters to set. In OpenCode, Skills are defined in SKILL.md files, allowing the LLM to load and execute reusable behaviours, which are more modular than plain prompts.

4. Agents

OpenCode provides two built‑in agent modes:

Build (primary agent) : the default agent with all tools enabled, suitable for everyday development that requires full file and system access.

Plan (primary agent) : a restricted agent designed for code analysis, suggestion, or planning without modifying the codebase.

The primary agent can switch to sub‑agents for specialised tasks. Sub‑agents run in independent sessions and are invoked either automatically by the primary agent or manually with @agent_name.

The most popular OpenCode plugin, Oh‑My‑OpenCode (OMO), expands this model to eleven specialised agents that work in parallel, coordinated by a scheduler, achieving efficiency far beyond a single agent.

Sisyphus – Recommended models: Claude Opus 4.6, Kimi‑K2.5, GLM‑5 – Role: primary orchestrator that plans tasks, delegates subtasks, and drives completion – Scenarios: complex development tasks, daily coding.

Hephaestus – Models: GPT‑5.3‑Codex, Claude Sonnet 4.6 – Role: autonomous deep worker that executes end‑to‑end without step‑by‑step guidance – Scenarios: large‑scale feature development, architecture refactoring.

Prometheus – Models: Claude Opus 4.6, Kimi‑K2.5, GLM‑5 – Role: strategic planner that conducts interview‑style requirement analysis and generates plans – Scenarios: multi‑day projects, new feature planning.

Oracle – Models: Claude Opus 4.6, GPT‑5.2, DeepSeek‑V3.2 – Role: architecture decision‑maker, complex debugging, code review – Scenarios: architecture design, post‑failure troubleshooting.

Librarian – Models: Gemini 3.1 Flash, Grok Code Fast, Qwen3‑Coder‑Next – Role: document query and OSS implementation search – Scenarios: third‑party library lookup, open‑source example discovery.

Explore – Models: Gemini 3.1 Flash, Claude Haiku 4.5, Qwen3‑Coder‑Next – Role: code‑base exploration and structural understanding – Scenarios: onboarding new codebases, pattern discovery.

Metis – Models: Claude Sonnet 4.6, GPT‑5 – Role: pre‑planning consultant that analyses hidden intents and failure points – Scenarios: scope clarification for complex tasks.

Momus – Models: Claude Sonnet 4.6, GPT‑5 – Role: plan reviewer that checks completeness and clarity – Scenarios: validation of generated plans.

Multimodal Looker – Models: Gemini 3.1 Pro, Claude Sonnet 4.6 – Role: image, PDF, chart analysis – Scenarios: screenshot analysis, document interpretation.

In summary, the fundamental concepts of agents, tools, MCP, rules, and skills are shared across OpenCode, Claude code, and other codex‑style systems. Moving from basic usage to a systematic, engineering‑grade workflow can be guided by the SDD methodology, which will be covered in future articles.