The short answer
| You want to… | Use | Where |
|---|---|---|
| Chat with Claude about your books, in the Claude app on your Mac or PC ("which invoices are overdue?", "draft a quote for…") | An API token, added to Claude as the Beeswax connector (MCP) | Connect to Claude & AI Assistants |
| Connect Zapier, n8n, a reporting tool or your own script | An API token | API Tokens |
| Assign tasks to an AI that works through them on its own, on one of your computers, and hands the result back for review | A desktop agent | Desktop Agents |
If you just want Claude to see your Beeswax data while you talk to it, you need an API token, not a desktop agent.
What "MCP" means here
MCP (Model Context Protocol) is the open standard Claude and other AI apps use to connect to outside services. The Beeswax MCP server is the connector you install in the AI app. It never has its own login: it always runs on a key you create in Beeswax.
- For Claude on your own computer, that key is an API token.
- For a desktop agent, that key is the agent's access key.
So both options below use the same connector. What differs is who the AI is working as and how it gets its work.
API token: the AI works as you
Create one under Account Settings → API Tokens, then paste it into the Beeswax connector in Claude (or into Zapier, n8n or your script).
- It works as you. Everything it does is done under your name, and it can never do more than your own role allows. The scopes you tick narrow that further.
- You're in the conversation. Claude acts when you ask it something, and you see the answer straight away.
- No seat. A token doesn't use a member seat.
- One token per tool, per computer. Name each one after what uses it, for example Claude: MacBook Pro, so you can revoke it on its own.
Start here: Connect to Claude & AI Assistants (MCP). It has a step-by-step video.
Desktop agent: the AI is a team member
Create one on the Desktop agents card, under API Tokens.
- It's its own team member. The agent appears in task dropdowns and member lists with a Desktop Agent label, and it uses a member seat. It can't sign in to the web app.
- It works from assigned tasks, unattended. You assign it a task like you would a person. An AI app on one of your computers (such as Claude Code or Codex) picks the task up on a schedule, does the work with the apps and files on that computer, and submits it for review.
- A person always approves. The agent can only submit work for review. Closing the task is always a person's decision.
- One agent, one computer, one key. For another computer, add another agent.
Start here: Desktop Agents.
Side by side
| API token | Desktop agent | |
|---|---|---|
| Who it works as | You | Itself, as a team member |
| How it gets work | You ask it, in a chat or a workflow | Tasks you assign it |
| Runs | When you use it | On a schedule, unattended |
| Uses a member seat | No | Yes |
| Work reviewed by a person | Up to you | Always |
| Connector | Beeswax MCP server (for AI apps) or the API | Beeswax MCP server |
Common mix-ups
- "I added a desktop agent but Claude can't see my data." A desktop agent's key only does what the agent was given and is meant for unattended task work. To chat with Claude about your books, create an API token and add it to Claude by following Connect to Claude & AI Assistants.
- "Do I need both?" Only if you want both. Most people start with an API token for Claude, and add a desktop agent later for recurring work they want done without them.
- "Is the token an AI subscription?" No. Tokens and agent keys only connect an app to Beeswax. The AI itself runs under your own Claude (or other) subscription.