AI Agents, Tools, Skills, and MCP Mean the Same Thing

I was talking to a lawyer about AI agents. He wanted to know which noun to buy: agents, Claude skills, MCP, or ChatGPT plugins. I told him they are mostly stickers on the same stack.

The only real difference is tool calls. Most frontier models can already do those.

Vendors keep renaming the parts. You do not need a new species every quarter. You need one layer that holds the systems you already run: case, email, docs, billing. That layer is the firm brain. FirmOps builds it and manages it so you are not shopping a new word every time the demos change.

The sticker map

Here is the plain map I use with owners. If a salesperson uses a different label next month, put it back on this list before you buy.

Prompt

The instruction. What you want done.

Skill

A saved prompt plus the recipe for how this firm always does that job. Claude skills live here. So do saved playbooks under other names.

Tool

A permissioned action the model can take. Look up a matter. Draft an email. Search the file.

MCP, connector, or plugin

How those tools get attached to the model. The plug. Claude says MCP. ChatGPT has said plugins, then GPTs, then connectors. Same idea.

Agent

A model that can use tools, often on a loop or a schedule. Not a new species.

Firm brain

The layer that holds this stack together across the systems the firm already runs. Read-first. Humans approve anything that leaves the firm.

That is the whole glossary. The rest of this post is the coffee version of each line.

A prompt is the instruction

A prompt is the ask. "What is stuck on this matter?" "Draft the provider follow-up from the file." That is it.

A sharper prompt can get a cleaner first draft. It cannot see Clio, Gmail, or Dropbox unless you attach a tool. It cannot follow your firm's recipe unless you save that recipe.

This is why "we wrote better prompts" rarely changes a PI shop. The instruction was never the bottleneck. The file was sitting in three systems, and the model could not see it.

Prompts vs agents is the same point. An agent is still running a prompt. It just has permission to act through tools while it works.

A skill is a saved prompt plus this firm's recipe

Claude calls these skills. ChatGPT has called similar things custom GPTs or saved instructions. Same job.

A skill is how this firm always does that work. Which letter template. Which tone. Which fields must be filled before a draft is ready. Which facts come from the matter, not from memory.

Ask the model to "draft a provider LOR" with no skill and you get a generic letter. Give it the skill that knows your template, your required facts, and where those facts live, and you get a first draft a case manager can actually review.

AI skills vs agents is not a fork in the road. A skill is the recipe. An agent is the model that can follow that recipe and call tools while it does. A skill with no tools is still a chatbot with a better memory. Useful. Not a firm brain.

A tool is a permissioned action

A tool is a thing the model is allowed to do. Look up a matter. Read the timeline. Search email. Find a Dropbox folder. Prepare a Gmail draft.

Permissioned is the word that matters. You do not hand the model the keys to the firm. You hand it a short menu. Read first. Draft second. A person approves before anything is sent, filed, or published.

The difference between AI tools and agents, for lawyers, is smaller than the brochures. A tool is the action. An agent is the model allowed to take that action. If the model cannot look up the file, you do not have a tool. You have a chat window.

MCP, connectors, and plugins are the plug

MCP is the Model Context Protocol. That is the Claude name for the plug. ChatGPT plugins for a law firm were the same job under an older sticker. Then the name became GPTs. Then connectors. Other vendors say integration or Zap.

It is how tools get attached to the model. Different vendors. Same idea.

If you want the operator walkthrough of Claude MCP in a staff window, read Claude Desktop plus MCP as a company brain. That post is how the plug works in practice. This post is the sticker guide so you do not treat MCP as a different product from a plugin.

When a vendor says their agent "has MCP," ask three questions. Which tools can it call? What can it write? Who approves the send? The protocol name is not the product.

An agent is a model that can use tools

An agent is not a new species. It is a model that can call tools. Sometimes it loops until the job is done. Sometimes it runs on a schedule. Sometimes a person starts it.

I told that lawyer they can think of agents as tools of MCP, or as a prompt of a skill. The only real difference is the tool calls. Most models can do those now.

So "AI agents for law firms" is not a shopping category. It is a description of a model that is allowed to act through a plug. If you buy an agent that cannot see your case system, email, and docs, you bought a chatbot with a louder name.

The Managed AI Agents page is the mechanism: design, deploy, monitor, tune. It is not a second buy. It is how the firm brain takes action under supervision.

The firm brain holds the stack together

The firm brain is not another sticker. It is the layer that keeps prompt, skill, tool, plug, and agent pointed at the same file.

It reads the systems you already run. Case. Email. Docs. Billing. It answers from those sources first. It drafts for review. A person approves anything that leaves the firm.

That is the Managed Firm Brain. It sits under the law firm operations umbrella. It does not replace Clio, Filevine, or SmartAdvocate. The buy is one supervised layer, not another tab.

At Conduit, from the COO seat, case launch went from 2.5 hours to 10 minutes. The firm has run 1,000+ cases through that path. A 47-step launch became 1 click. That did not happen because we bought a better noun. It happened because the work sat under one supervised layer.

Stop buying a new noun every quarter

If you googled AI agents, MCP, Claude skills, or ChatGPT plugins, you were not looking for a glossary. You were looking for a way to stop retyping the same facts.

Do not hire a stack of new words. We build and manage the firm brain so the words stay boring.

Bring one recurring bottleneck to a 30-minute live demo. Email hello@firmops.io. We will map it to prompt, skill, tool, plug, and approval. Then we will tell you whether it is a first build.

Questions owners actually ask

What are AI agents for law firms?

A model that can use tools on your file, often on a loop or a schedule. Not a new kind of software. Most frontier models can already do tool calls. The useful version reads the systems you already run and waits for a human before anything leaves the firm.

What is MCP for lawyers?

MCP is a plug. It attaches permissioned tools to a model, the way ChatGPT plugins or connectors do. Claude MCP is that plug with Claude's name on it. The value is which tools it can call and who approves the write, not the acronym.

What is the difference between AI tools and agents?

A tool is the action. An agent is the model allowed to take that action. Lawyers get sold them as separate products. They are two layers of the same stack.

What are Claude skills, and how do they differ from agents?

A Claude skill is a saved prompt plus this firm's recipe for a job. An agent can follow that skill and call tools while it works. Keep the recipe and the plug together. Do not buy a new noun for each.

Are ChatGPT plugins the same as MCP?

Same job, different sticker. Plugins, GPTs, connectors, and MCP all attach tools to a model. Ask what the model can read, what it can draft, and who releases the send.