Quick answer
A language model is not a product. It is a seat. Chat when a person is in the loop. An IDE or terminal agent when the artefact is code. A workflow node when the same job runs without anyone prompting. An API when a customer hits your app. Most teams buy the chat and then paste. That is fine until the paste is the business. Then you wanted AI automation or a feature, not a better tab.
Key takeaways
- Name the seat before you name the vendor.
- Claude Pro is $20 / month ($17 annual) for a person. API usage is the product. n8n Cloud Starter is €20 / month yearly for 2,500 executions (Sep 2026 list). Do not confuse those invoices.
- If nobody needs to chat, do not put a chat in the middle.
- An agent that “does the business” is usually a workflow with one model step.
/01
Five seats
- Chat
- A person asks, a person checks. Briefs, rewrites, a first pass. ChatGPT, Claude, or Gemini. See ChatGPT vs Claude vs Gemini.
- IDE agent
- The artefact is a repo you live in. Cursor, Copilot. You still click the diff. See AI coding tools.
- Terminal agent
- A named slice, a long loop, a checkout. Claude Code. Same review rule.
- Workflow node
- The job is the same every Tuesday. Classify, extract, draft, stop if unsure. Claude as a step in n8n.
- Product API
- A user of yours hits a feature. You own the prompt, the log, the fallback. That is AI development, not a Plus seat.
MCP is not a sixth seat. It is how a chat or an agent reaches a tool. Read MCP explained if the problem is copy-paste, not “which model.”
/02
A short picker
- A person must see it before it leaves the building → chat
- The output is a pull request → IDE or terminal
- The input arrives on a webhook → workflow node
- A customer waits on the response → product API, with a timeout and a human path
The usual waste
Paying for Claude, ChatGPT, and Gemini, then pasting the same paragraph through all three. Pick one chat. Spend the rest of the money on the seat that acts.
/03
AnswersFAQ
faq.mdx
01
Is an agent a different seat?
An agent is a model that takes the next step. It still sits in a chat, a workflow, or your API. Name that seat first.02
Can n8n replace the chat?
If nobody needs to prompt, yes. A digest that posts to Slack does not need Claude.ai open.03
Do we need MCP for a product feature?
Usually no. MCP is for a model that must reach existing tools. A feature you own can call your own API.04
Where should a small team start?
One chat seat. One workflow if a human is the integration. Add an API when a customer will feel the answer.Have something to build? Let’s ship it.
Book a call