AI Tools

The words that come with AI agents — tool, loop, harness, skill, plugin — explained in the order they build on each other, then a short list of tools worth using for code, testing, images, video and building your own, with which one to pick when.

How the pieces fit

A model only turns text into text. Everything that makes it do things is wrapped around it, layer by layer — each term below is one of those layers.

Pluginpackages skills, hooks, subagents, MCP config

Harnessruns the loop, owns permissions and context

Agent loopmodel → tool call → result → model …

Model + Toolsthe model decides; tools act

Model (LLM)

The thing that predicts the next token.

A large language model takes text (and often images) in and produces text out. On its own it cannot read files, browse or remember anything between calls — everything it knows about your task has to be in the request.

e.g. Claude, GPT, Gemini, Llama, DeepSeek, Qwen.

Not the app. ChatGPT and Claude.ai are products built around a model; the model is the engine inside.

Prompt & context window

Everything the model can see in one request.

The system prompt, the conversation so far, file contents, tool results — all of it is packed into the context window, which has a fixed size measured in tokens. When it fills up, something has to be summarised or dropped.

e.g. A coding agent pasting three source files and your question into one request.

Not memory. Nothing survives to the next request unless the app sends it again.

Tool (function calling)

A function the model may ask the app to run.

The app describes each tool — a name, a description and a JSON schema for its arguments. The model replies with "call read_file with path=…" instead of prose; the app runs it and sends the result back. The model never executes anything itself.

e.g. read_file, run_shell, web_search, query_database.

Not an "AI tool" in the product sense. In agent vocabulary a tool is one callable function.

Agent loop

Call the model, run its tools, repeat until done.

The loop is what turns a chat model into something that does work: send context → model asks for a tool → run it → append the result → call the model again. It stops when the model answers without asking for a tool, or hits a limit.

e.g. Fix a failing test: read the test → run it → read the source → edit → run again → report.

Not a cron-style loop. "/loop" in Claude Code re-runs a whole prompt on a schedule; the agent loop is the inner cycle of one task.

AI agent

A model plus tools plus a loop, pointed at a goal.

An agent decides its own next step rather than following a fixed script. The same model becomes a different agent depending on its instructions and which tools it is given — a read-only research agent and a code-editing agent can share one model.

e.g. A support agent that looks up an order, checks the refund policy and drafts a reply.

Not a workflow. A workflow's steps are decided by the author in advance; an agent's are decided by the model at run time.

Harness

The program wrapped around the model that runs the loop.

The harness owns everything the model cannot: the loop itself, the built-in tools, permissions and sandboxing, what goes into the context and when to compact it, and the UI. Two products on the same model can behave very differently because their harnesses differ.

e.g. Claude Code, Cursor's agent, OpenAI Codex CLI, Gemini CLI.

Not the model. Swapping the model changes how well it reasons; swapping the harness changes what it is able to do.

MCP server

A standard plug for giving any harness new tools.

The Model Context Protocol lets a separate process expose tools (and resources and prompts) to any harness that speaks MCP. Write a GitHub or database connector once and it works in Claude Code, Cursor, VS Code and others.

e.g. A MySQL MCP server exposing list_tables and execute_query.

Not a skill. MCP adds new actions the agent can take; a skill adds know-how about how to take them.

Skill

A folder of instructions the agent loads only when relevant.

A skill is a SKILL.md file (plus optional scripts and references) with a name and a one-line description. Only the description sits in context; the full instructions are read when a task matches, so you can install many without filling the context window.

e.g. A deploy-prod skill describing how to tag a release and push it.

Not a tool. A skill contains no new capability — it teaches the agent a procedure using the tools it already has.

Subagent

A second agent the main one hands a sub-task to.

The subagent runs its own loop in a fresh context and returns only a summary. It keeps big searches and file dumps out of the main conversation, and lets independent jobs run in parallel.

e.g. An explore agent that sweeps 200 files and reports back the three that matter.

Not a separate product. It is the same harness starting another loop with narrower instructions and tools.

Hook

Your own script the harness runs at a fixed moment.

Hooks fire on events — before a tool runs, after a file is edited, when the agent stops. They are deterministic: the harness runs them every time, rather than hoping the model remembers an instruction.

e.g. Run the formatter after every file edit; block rm -rf before it executes.

Not a prompt instruction. "Always run lint" in a prompt is a request; a hook is a guarantee.

Plugin

An installable bundle of skills, commands, agents, hooks and MCP servers.

A plugin is the packaging layer: one install that brings a coherent set of the pieces above, usually from a marketplace or a git repo. It exists so a team can share a working setup instead of copying files around.

e.g. A code-review plugin shipping a /review command, a reviewer subagent and a lint hook.

Not a capability of its own. Everything a plugin does is done by the skills, hooks and servers inside it.

RAG

Retrieve relevant documents, then let the model answer from them.

Retrieval-augmented generation searches your own data — usually via embeddings in a vector database — and puts the best matches into the prompt. It is how a model answers about documents it was never trained on.

e.g. A help-centre bot that finds the three closest articles before answering.

Not fine-tuning. RAG changes what the model reads per request; fine-tuning changes the model's weights.

Tool vs MCP vs skill vs hook vs plugin

Piece	What it adds	When it runs	What it looks like
Tool	One new action	The model chooses to call it	A function + JSON schema
MCP server	A set of tools, reusable across harnesses	The model calls one of its tools	A separate process speaking MCP
Skill	Know-how, not new actions	The task matches its description	A folder with SKILL.md
Hook	A guaranteed side effect or check	Always, on a fixed event	A shell command in settings
Plugin	All of the above, installed together	Whenever its parts do	A package from a marketplace

Recommended tools

A short list per job rather than every product on the market. Pricing and model versions change monthly, so check the vendor’s site before committing.

Reading GitHub stars and weekly downloads…

35 of 35 tools

Coding agents

Harnesses that read, edit and run code in your repo. The model matters, but the harness decides how much of your project it can see and do.

Claude Code

Anthropic

Paid★ Good

Autonomous multi-file work from the terminal or IDE

Terminal-first agent with skills, hooks, subagents, plugins and MCP. Runs in the CLI, VS Code, JetBrains, desktop and web.

Which coding agent?
Claude Code	You want to hand over whole tasks and customise the agent with skills, hooks and subagents.
Cursor	You mostly write the code yourself and want fast completions and inline edits in an editor.
GitHub Copilot	Your company already pays for GitHub and you want agents working from issues and PRs.
Codex / Gemini CLI	You prefer OpenAI or Google models, or want an open-source CLI you can read.
Aider	You want every change as a git commit and the freedom to use any model, including local ones.

Which image tool?
Midjourney	It should look beautiful and you are flexible on the details.
ChatGPT / Gemini	It must follow the brief exactly, contain readable text, or edit an existing photo.
Ideogram / Recraft	The output is a logo, poster, icon or SVG.
Adobe Firefly	It is commercial work and you need clear usage rights.
FLUX + ComfyUI	You want it free, private and fully controllable on your own GPU.

Which video tool?
Veo / Sora	You need the most realistic clip, with sound, from a text prompt.
Kling / Luma	You are animating a still image or need many cheap iterations.
Runway	You are editing or restyling real footage and need fine camera and motion control.
HeyGen	Someone needs to talk to camera — explainers, training, translated versions.

LangChain, LangGraph or something else?
LangChain	A straightforward agent or chain where swapping providers and integrations matters most.
LangGraph	Long-running, branching agents that need saved state, retries or a human approval step.
LlamaIndex	The hard part is the documents — parsing, chunking and retrieving them well.
Vercel AI SDK	You are building a chat or AI feature into a TypeScript web app.
Claude / OpenAI Agents SDK	You are committed to one vendor and want its agent loop without building your own.

AI Tools

How the pieces fit

Tool vs MCP vs skill vs hook vs plugin

Recommended tools

Coding agents

Testing agents

Image

Video

LLM frameworks

Visual builders

Run locally

Model APIs